diff --git "a/trainer_state.json" "b/trainer_state.json"
new file mode 100644--- /dev/null
+++ "b/trainer_state.json"
@@ -0,0 +1,42033 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.875,
+  "eval_steps": 2000,
+  "global_step": 35000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.00025,
+      "grad_norm": 31.5,
+      "learning_rate": 0.0001,
+      "loss": 7.633,
+      "loss/crossentropy": 2.065455098450184,
+      "loss/hidden": 3.476953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20220321230590343,
+      "step": 10
+    },
+    {
+      "epoch": 0.0005,
+      "grad_norm": 35.0,
+      "grad_norm_var": 2.6895182291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4618,
+      "loss/crossentropy": 1.9399560801684856,
+      "loss/hidden": 3.394140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19191570337861777,
+      "step": 20
+    },
+    {
+      "epoch": 0.00075,
+      "grad_norm": 37.5,
+      "grad_norm_var": 6.579622395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5972,
+      "loss/crossentropy": 2.130601316690445,
+      "loss/hidden": 3.38984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20188977513462306,
+      "step": 30
+    },
+    {
+      "epoch": 0.001,
+      "grad_norm": 33.5,
+      "grad_norm_var": 6.253125,
+      "learning_rate": 0.0001,
+      "loss": 7.5917,
+      "loss/crossentropy": 2.2571407079696657,
+      "loss/hidden": 3.422265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19847887996584176,
+      "step": 40
+    },
+    {
+      "epoch": 0.00125,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.1619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.6054,
+      "loss/crossentropy": 2.1717565625905992,
+      "loss/hidden": 3.43359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20264342725276946,
+      "step": 50
+    },
+    {
+      "epoch": 0.0015,
+      "grad_norm": 35.5,
+      "grad_norm_var": 15.786393229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5513,
+      "loss/crossentropy": 2.070718301087618,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19855907820165158,
+      "step": 60
+    },
+    {
+      "epoch": 0.00175,
+      "grad_norm": 31.0,
+      "grad_norm_var": 12.4625,
+      "learning_rate": 0.0001,
+      "loss": 7.5447,
+      "loss/crossentropy": 2.118075390160084,
+      "loss/hidden": 3.473828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20283062420785428,
+      "step": 70
+    },
+    {
+      "epoch": 0.002,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.2643229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.468,
+      "loss/crossentropy": 2.0006178975105287,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18958428762853147,
+      "step": 80
+    },
+    {
+      "epoch": 0.00225,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.470572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5061,
+      "loss/crossentropy": 1.9605075903236866,
+      "loss/hidden": 3.54375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20559987109154462,
+      "step": 90
+    },
+    {
+      "epoch": 0.0025,
+      "grad_norm": 31.125,
+      "grad_norm_var": 6.763541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4928,
+      "loss/crossentropy": 2.1205389350652695,
+      "loss/hidden": 3.44140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19496036488562823,
+      "step": 100
+    },
+    {
+      "epoch": 0.00275,
+      "grad_norm": 31.0,
+      "grad_norm_var": 6.1509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.595,
+      "loss/crossentropy": 2.1240097641944886,
+      "loss/hidden": 3.43671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19564666803926228,
+      "step": 110
+    },
+    {
+      "epoch": 0.003,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.348893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5329,
+      "loss/crossentropy": 2.175096944719553,
+      "loss/hidden": 3.41796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21303062327206135,
+      "step": 120
+    },
+    {
+      "epoch": 0.00325,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.8541666666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5536,
+      "loss/crossentropy": 2.1472502022981645,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18929538186639547,
+      "step": 130
+    },
+    {
+      "epoch": 0.0035,
+      "grad_norm": 29.375,
+      "grad_norm_var": 29.683268229166668,
+      "learning_rate": 0.0001,
+      "loss": 7.5191,
+      "loss/crossentropy": 2.015011890232563,
+      "loss/hidden": 3.44296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20328481420874595,
+      "step": 140
+    },
+    {
+      "epoch": 0.00375,
+      "grad_norm": 28.75,
+      "grad_norm_var": 28.74765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4158,
+      "loss/crossentropy": 1.9774167470633983,
+      "loss/hidden": 3.43515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19464388117194176,
+      "step": 150
+    },
+    {
+      "epoch": 0.004,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.3635416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6354,
+      "loss/crossentropy": 2.320629420876503,
+      "loss/hidden": 3.418359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20745602920651435,
+      "step": 160
+    },
+    {
+      "epoch": 0.00425,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.0270182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4137,
+      "loss/crossentropy": 1.900385806709528,
+      "loss/hidden": 3.345703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16769229620695114,
+      "step": 170
+    },
+    {
+      "epoch": 0.0045,
+      "grad_norm": 31.25,
+      "grad_norm_var": 0.9833333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5763,
+      "loss/crossentropy": 2.129625543951988,
+      "loss/hidden": 3.5171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2102549459785223,
+      "step": 180
+    },
+    {
+      "epoch": 0.00475,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.05390625,
+      "learning_rate": 0.0001,
+      "loss": 7.6166,
+      "loss/crossentropy": 2.1552532628178596,
+      "loss/hidden": 3.469140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2250068686902523,
+      "step": 190
+    },
+    {
+      "epoch": 0.005,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.8375,
+      "learning_rate": 0.0001,
+      "loss": 7.5745,
+      "loss/crossentropy": 1.9441482461988926,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.195942450594157,
+      "step": 200
+    },
+    {
+      "epoch": 0.00525,
+      "grad_norm": 32.5,
+      "grad_norm_var": 18.396875,
+      "learning_rate": 0.0001,
+      "loss": 7.5292,
+      "loss/crossentropy": 1.9941987417638303,
+      "loss/hidden": 3.394140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18264975901693106,
+      "step": 210
+    },
+    {
+      "epoch": 0.0055,
+      "grad_norm": 31.75,
+      "grad_norm_var": 20.736393229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4899,
+      "loss/crossentropy": 2.0191620789468288,
+      "loss/hidden": 3.355078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18100650198757648,
+      "step": 220
+    },
+    {
+      "epoch": 0.00575,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.342643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5199,
+      "loss/crossentropy": 2.001779730618,
+      "loss/hidden": 3.32109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17959208656102418,
+      "step": 230
+    },
+    {
+      "epoch": 0.006,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.271875,
+      "learning_rate": 0.0001,
+      "loss": 7.6842,
+      "loss/crossentropy": 2.1846971333026888,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2059234745800495,
+      "step": 240
+    },
+    {
+      "epoch": 0.00625,
+      "grad_norm": 29.5,
+      "grad_norm_var": 5.688541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5196,
+      "loss/crossentropy": 2.174124576151371,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20000722594559192,
+      "step": 250
+    },
+    {
+      "epoch": 0.0065,
+      "grad_norm": 28.75,
+      "grad_norm_var": 1.9572265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3875,
+      "loss/crossentropy": 1.9285166233778,
+      "loss/hidden": 3.396875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18449910767376423,
+      "step": 260
+    },
+    {
+      "epoch": 0.00675,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.0999348958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5877,
+      "loss/crossentropy": 2.0323276594281197,
+      "loss/hidden": 3.37890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19395631980150937,
+      "step": 270
+    },
+    {
+      "epoch": 0.007,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.15390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5791,
+      "loss/crossentropy": 2.126656140387058,
+      "loss/hidden": 3.496875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21661139875650406,
+      "step": 280
+    },
+    {
+      "epoch": 0.00725,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.193489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5587,
+      "loss/crossentropy": 2.200097793340683,
+      "loss/hidden": 3.529296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21046234332025052,
+      "step": 290
+    },
+    {
+      "epoch": 0.0075,
+      "grad_norm": 26.75,
+      "grad_norm_var": 4.27265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5404,
+      "loss/crossentropy": 2.1184144005179406,
+      "loss/hidden": 3.487890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20949590150266886,
+      "step": 300
+    },
+    {
+      "epoch": 0.00775,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.3643229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5628,
+      "loss/crossentropy": 1.9984030593186617,
+      "loss/hidden": 3.453515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18789457948878407,
+      "step": 310
+    },
+    {
+      "epoch": 0.008,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.5645182291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5695,
+      "loss/crossentropy": 2.143594169616699,
+      "loss/hidden": 3.42421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19360470157116652,
+      "step": 320
+    },
+    {
+      "epoch": 0.00825,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.8749348958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3627,
+      "loss/crossentropy": 2.1077703177928924,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19771252572536469,
+      "step": 330
+    },
+    {
+      "epoch": 0.0085,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.5978515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4192,
+      "loss/crossentropy": 2.0583472289144993,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20189273860305548,
+      "step": 340
+    },
+    {
+      "epoch": 0.00875,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.2872395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5432,
+      "loss/crossentropy": 2.0804511278867723,
+      "loss/hidden": 3.38828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19735569059848784,
+      "step": 350
+    },
+    {
+      "epoch": 0.009,
+      "grad_norm": 30.5,
+      "grad_norm_var": 18.731184895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4948,
+      "loss/crossentropy": 2.0466629534959795,
+      "loss/hidden": 3.315234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18366040643304588,
+      "step": 360
+    },
+    {
+      "epoch": 0.00925,
+      "grad_norm": 30.875,
+      "grad_norm_var": 25.9916015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5081,
+      "loss/crossentropy": 1.9005662694573402,
+      "loss/hidden": 3.501171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1900689721107483,
+      "step": 370
+    },
+    {
+      "epoch": 0.0095,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.451041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4305,
+      "loss/crossentropy": 2.0674299761652946,
+      "loss/hidden": 3.517578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21062961965799332,
+      "step": 380
+    },
+    {
+      "epoch": 0.00975,
+      "grad_norm": 31.25,
+      "grad_norm_var": 5.645247395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5168,
+      "loss/crossentropy": 2.0279919117689134,
+      "loss/hidden": 3.503125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20519332773983479,
+      "step": 390
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 31.125,
+      "grad_norm_var": 5.928125,
+      "learning_rate": 0.0001,
+      "loss": 7.4985,
+      "loss/crossentropy": 2.0427632443606853,
+      "loss/hidden": 3.53125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20287631042301654,
+      "step": 400
+    },
+    {
+      "epoch": 0.01025,
+      "grad_norm": 38.5,
+      "grad_norm_var": 438.43515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5633,
+      "loss/crossentropy": 2.199043881893158,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21130343191325665,
+      "step": 410
+    },
+    {
+      "epoch": 0.0105,
+      "grad_norm": 30.875,
+      "grad_norm_var": 43.14140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4835,
+      "loss/crossentropy": 1.9102243572473525,
+      "loss/hidden": 3.42578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1895731385797262,
+      "step": 420
+    },
+    {
+      "epoch": 0.01075,
+      "grad_norm": 31.75,
+      "grad_norm_var": 5.658268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3897,
+      "loss/crossentropy": 2.159160128980875,
+      "loss/hidden": 3.464453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20280379485338926,
+      "step": 430
+    },
+    {
+      "epoch": 0.011,
+      "grad_norm": 28.375,
+      "grad_norm_var": 16.3375,
+      "learning_rate": 0.0001,
+      "loss": 7.5463,
+      "loss/crossentropy": 2.1217672407627104,
+      "loss/hidden": 3.545703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.23856931366026402,
+      "step": 440
+    },
+    {
+      "epoch": 0.01125,
+      "grad_norm": 30.5,
+      "grad_norm_var": 17.098372395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5225,
+      "loss/crossentropy": 1.969854873791337,
+      "loss/hidden": 3.430078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19548849146813155,
+      "step": 450
+    },
+    {
+      "epoch": 0.0115,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.5677083333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5046,
+      "loss/crossentropy": 2.121321603655815,
+      "loss/hidden": 3.476171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19364523217082025,
+      "step": 460
+    },
+    {
+      "epoch": 0.01175,
+      "grad_norm": 32.25,
+      "grad_norm_var": 8.585416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4558,
+      "loss/crossentropy": 1.9360710382461548,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1893781816586852,
+      "step": 470
+    },
+    {
+      "epoch": 0.012,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.417122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.531,
+      "loss/crossentropy": 2.082458943128586,
+      "loss/hidden": 3.471875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2220946006476879,
+      "step": 480
+    },
+    {
+      "epoch": 0.01225,
+      "grad_norm": 31.0,
+      "grad_norm_var": 48.96640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5651,
+      "loss/crossentropy": 2.1382531195878984,
+      "loss/hidden": 3.480078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20847559962421655,
+      "step": 490
+    },
+    {
+      "epoch": 0.0125,
+      "grad_norm": 29.875,
+      "grad_norm_var": 49.2666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5679,
+      "loss/crossentropy": 2.0875915244221686,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1850985599681735,
+      "step": 500
+    },
+    {
+      "epoch": 0.01275,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.45,
+      "learning_rate": 0.0001,
+      "loss": 7.5263,
+      "loss/crossentropy": 2.182442346215248,
+      "loss/hidden": 3.446484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19555890336632728,
+      "step": 510
+    },
+    {
+      "epoch": 0.013,
+      "grad_norm": 34.0,
+      "grad_norm_var": 1.6931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5209,
+      "loss/crossentropy": 1.9812136888504028,
+      "loss/hidden": 3.4921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1965757070109248,
+      "step": 520
+    },
+    {
+      "epoch": 0.01325,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.101822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6059,
+      "loss/crossentropy": 2.0372241511940956,
+      "loss/hidden": 3.564453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.204646560549736,
+      "step": 530
+    },
+    {
+      "epoch": 0.0135,
+      "grad_norm": 29.125,
+      "grad_norm_var": 20.071875,
+      "learning_rate": 0.0001,
+      "loss": 7.5725,
+      "loss/crossentropy": 2.155761349201202,
+      "loss/hidden": 3.4125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19602423422038556,
+      "step": 540
+    },
+    {
+      "epoch": 0.01375,
+      "grad_norm": 29.125,
+      "grad_norm_var": 20.506705729166665,
+      "learning_rate": 0.0001,
+      "loss": 7.5842,
+      "loss/crossentropy": 1.8869566857814788,
+      "loss/hidden": 3.437890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20957522764801978,
+      "step": 550
+    },
+    {
+      "epoch": 0.014,
+      "grad_norm": 30.625,
+      "grad_norm_var": 10.025455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4975,
+      "loss/crossentropy": 2.0370677679777147,
+      "loss/hidden": 3.361328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19026046600192786,
+      "step": 560
+    },
+    {
+      "epoch": 0.01425,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.2270833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5688,
+      "loss/crossentropy": 2.1931444257497787,
+      "loss/hidden": 3.415234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2036376902833581,
+      "step": 570
+    },
+    {
+      "epoch": 0.0145,
+      "grad_norm": 35.0,
+      "grad_norm_var": 3.5681640625,
+      "learning_rate": 0.0001,
+      "loss": 7.478,
+      "loss/crossentropy": 2.061052493005991,
+      "loss/hidden": 3.478125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2282864760607481,
+      "step": 580
+    },
+    {
+      "epoch": 0.01475,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.8705729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5957,
+      "loss/crossentropy": 2.0078392371535303,
+      "loss/hidden": 3.45,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19647251404821872,
+      "step": 590
+    },
+    {
+      "epoch": 0.015,
+      "grad_norm": 30.25,
+      "grad_norm_var": 31.449934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5096,
+      "loss/crossentropy": 2.0417068414390087,
+      "loss/hidden": 3.423046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19782953998073935,
+      "step": 600
+    },
+    {
+      "epoch": 0.01525,
+      "grad_norm": 30.5,
+      "grad_norm_var": 26.253059895833335,
+      "learning_rate": 0.0001,
+      "loss": 7.5368,
+      "loss/crossentropy": 2.1738049775362014,
+      "loss/hidden": 3.409765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1996332859620452,
+      "step": 610
+    },
+    {
+      "epoch": 0.0155,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.334375,
+      "learning_rate": 0.0001,
+      "loss": 7.4868,
+      "loss/crossentropy": 1.7587297886610032,
+      "loss/hidden": 3.475390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18938990794122218,
+      "step": 620
+    },
+    {
+      "epoch": 0.01575,
+      "grad_norm": 29.25,
+      "grad_norm_var": 27.393684895833335,
+      "learning_rate": 0.0001,
+      "loss": 7.4833,
+      "loss/crossentropy": 1.9551145888864994,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20075901364907622,
+      "step": 630
+    },
+    {
+      "epoch": 0.016,
+      "grad_norm": 29.75,
+      "grad_norm_var": 29.6947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4608,
+      "loss/crossentropy": 2.128718316555023,
+      "loss/hidden": 3.3625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19077460393309592,
+      "step": 640
+    },
+    {
+      "epoch": 0.01625,
+      "grad_norm": 29.75,
+      "grad_norm_var": 27.322330729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.6033,
+      "loss/crossentropy": 1.9678708665072917,
+      "loss/hidden": 3.413671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18875791020691396,
+      "step": 650
+    },
+    {
+      "epoch": 0.0165,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.129622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3873,
+      "loss/crossentropy": 1.9582339562475681,
+      "loss/hidden": 3.34765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18309127148240806,
+      "step": 660
+    },
+    {
+      "epoch": 0.01675,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.7009765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4913,
+      "loss/crossentropy": 2.0773802563548087,
+      "loss/hidden": 3.505078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20910798981785775,
+      "step": 670
+    },
+    {
+      "epoch": 0.017,
+      "grad_norm": 34.0,
+      "grad_norm_var": 3.3854166666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4847,
+      "loss/crossentropy": 2.12913373708725,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.201920267008245,
+      "step": 680
+    },
+    {
+      "epoch": 0.01725,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.7176432291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5065,
+      "loss/crossentropy": 1.9141538538038732,
+      "loss/hidden": 3.44921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1841401271522045,
+      "step": 690
+    },
+    {
+      "epoch": 0.0175,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.6374348958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5897,
+      "loss/crossentropy": 2.207232800126076,
+      "loss/hidden": 3.399609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21376523859798907,
+      "step": 700
+    },
+    {
+      "epoch": 0.01775,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.3655598958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5075,
+      "loss/crossentropy": 2.03845998942852,
+      "loss/hidden": 3.41953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1920805646572262,
+      "step": 710
+    },
+    {
+      "epoch": 0.018,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.3893229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4669,
+      "loss/crossentropy": 2.054341807588935,
+      "loss/hidden": 3.489453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19716067584231495,
+      "step": 720
+    },
+    {
+      "epoch": 0.01825,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.54140625,
+      "learning_rate": 0.0001,
+      "loss": 7.517,
+      "loss/crossentropy": 2.2111608639359472,
+      "loss/hidden": 3.409765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20262118335813284,
+      "step": 730
+    },
+    {
+      "epoch": 0.0185,
+      "grad_norm": 29.125,
+      "grad_norm_var": 4.692122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4784,
+      "loss/crossentropy": 2.0551758617162705,
+      "loss/hidden": 3.446875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20378697756677866,
+      "step": 740
+    },
+    {
+      "epoch": 0.01875,
+      "grad_norm": 33.0,
+      "grad_norm_var": 4.295572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4016,
+      "loss/crossentropy": 2.128055375814438,
+      "loss/hidden": 3.3953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19904747987166047,
+      "step": 750
+    },
+    {
+      "epoch": 0.019,
+      "grad_norm": 6106906624.0,
+      "grad_norm_var": 2.3308942582349476e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4633,
+      "loss/crossentropy": 2.248567137122154,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19723597317934036,
+      "step": 760
+    },
+    {
+      "epoch": 0.01925,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.330894258158611e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4542,
+      "loss/crossentropy": 2.132212319970131,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18174959290772677,
+      "step": 770
+    },
+    {
+      "epoch": 0.0195,
+      "grad_norm": 36.5,
+      "grad_norm_var": 4.833333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.465,
+      "loss/crossentropy": 2.046277052164078,
+      "loss/hidden": 3.491015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21161840241402388,
+      "step": 780
+    },
+    {
+      "epoch": 0.01975,
+      "grad_norm": 32.75,
+      "grad_norm_var": 5.137434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4171,
+      "loss/crossentropy": 2.058088332414627,
+      "loss/hidden": 3.315234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1815673651173711,
+      "step": 790
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 30.125,
+      "grad_norm_var": 12.37265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4153,
+      "loss/crossentropy": 2.064726157486439,
+      "loss/hidden": 3.515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19402222614735365,
+      "step": 800
+    },
+    {
+      "epoch": 0.02025,
+      "grad_norm": 32.0,
+      "grad_norm_var": 12.240625,
+      "learning_rate": 0.0001,
+      "loss": 7.3739,
+      "loss/crossentropy": 2.0926051691174505,
+      "loss/hidden": 3.476953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21017331834882497,
+      "step": 810
+    },
+    {
+      "epoch": 0.0205,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.6853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.409,
+      "loss/crossentropy": 2.016859006881714,
+      "loss/hidden": 3.436328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20363395065069198,
+      "step": 820
+    },
+    {
+      "epoch": 0.02075,
+      "grad_norm": 34.0,
+      "grad_norm_var": 278.1108723958333,
+      "learning_rate": 0.0001,
+      "loss": 7.6725,
+      "loss/crossentropy": 2.03957434669137,
+      "loss/hidden": 3.4625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19866096526384353,
+      "step": 830
+    },
+    {
+      "epoch": 0.021,
+      "grad_norm": 35.75,
+      "grad_norm_var": 281.2239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4058,
+      "loss/crossentropy": 2.1190530106425287,
+      "loss/hidden": 3.41796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19663097113370895,
+      "step": 840
+    },
+    {
+      "epoch": 0.02125,
+      "grad_norm": 32.25,
+      "grad_norm_var": 4.044791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4687,
+      "loss/crossentropy": 2.1552326917648315,
+      "loss/hidden": 3.41796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19604418501257898,
+      "step": 850
+    },
+    {
+      "epoch": 0.0215,
+      "grad_norm": 37.25,
+      "grad_norm_var": 2.7587362193217157e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5552,
+      "loss/crossentropy": 2.1164004117250443,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19724889248609542,
+      "step": 860
+    },
+    {
+      "epoch": 0.02175,
+      "grad_norm": 35.25,
+      "grad_norm_var": 2.758736219342478e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5021,
+      "loss/crossentropy": 2.036998500674963,
+      "loss/hidden": 3.298828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18320635841228067,
+      "step": 870
+    },
+    {
+      "epoch": 0.022,
+      "grad_norm": 37.0,
+      "grad_norm_var": 16.9541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5059,
+      "loss/crossentropy": 1.9707016140222549,
+      "loss/hidden": 3.36328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20436920877546072,
+      "step": 880
+    },
+    {
+      "epoch": 0.02225,
+      "grad_norm": 31.375,
+      "grad_norm_var": 30.538541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4935,
+      "loss/crossentropy": 2.206394499540329,
+      "loss/hidden": 3.366015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20495780408382416,
+      "step": 890
+    },
+    {
+      "epoch": 0.0225,
+      "grad_norm": 29.875,
+      "grad_norm_var": 28.020833333333332,
+      "learning_rate": 0.0001,
+      "loss": 7.4823,
+      "loss/crossentropy": 2.091763325035572,
+      "loss/hidden": 3.43828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20592593550682067,
+      "step": 900
+    },
+    {
+      "epoch": 0.02275,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.5645182291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.422,
+      "loss/crossentropy": 1.9740761511027813,
+      "loss/hidden": 3.494921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2015986293554306,
+      "step": 910
+    },
+    {
+      "epoch": 0.023,
+      "grad_norm": 32.0,
+      "grad_norm_var": 56.256184895833336,
+      "learning_rate": 0.0001,
+      "loss": 7.4528,
+      "loss/crossentropy": 2.030415116250515,
+      "loss/hidden": 3.205078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1614784031175077,
+      "step": 920
+    },
+    {
+      "epoch": 0.02325,
+      "grad_norm": 30.0,
+      "grad_norm_var": 57.1619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3713,
+      "loss/crossentropy": 2.0250086903572084,
+      "loss/hidden": 3.455859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19023355115205048,
+      "step": 930
+    },
+    {
+      "epoch": 0.0235,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.3830729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5277,
+      "loss/crossentropy": 2.222324788570404,
+      "loss/hidden": 3.366796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19078677501529456,
+      "step": 940
+    },
+    {
+      "epoch": 0.02375,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.1455729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5086,
+      "loss/crossentropy": 2.1299516543745995,
+      "loss/hidden": 3.49921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21310927756130696,
+      "step": 950
+    },
+    {
+      "epoch": 0.024,
+      "grad_norm": 29.875,
+      "grad_norm_var": 8.883072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5579,
+      "loss/crossentropy": 2.0535727672278883,
+      "loss/hidden": 3.43828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18507701791822911,
+      "step": 960
+    },
+    {
+      "epoch": 0.02425,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.5916015625,
+      "learning_rate": 0.0001,
+      "loss": 7.537,
+      "loss/crossentropy": 2.1785535484552385,
+      "loss/hidden": 3.309765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1955953363329172,
+      "step": 970
+    },
+    {
+      "epoch": 0.0245,
+      "grad_norm": 36.5,
+      "grad_norm_var": 6.852083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5091,
+      "loss/crossentropy": 2.0967498391866686,
+      "loss/hidden": 3.43515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2146583067253232,
+      "step": 980
+    },
+    {
+      "epoch": 0.02475,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.325455729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5901,
+      "loss/crossentropy": 2.1134474128484726,
+      "loss/hidden": 3.3953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19056662563234567,
+      "step": 990
+    },
+    {
+      "epoch": 0.025,
+      "grad_norm": 42.0,
+      "grad_norm_var": 4.1552039405313587e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.6082,
+      "loss/crossentropy": 2.0916516482830048,
+      "loss/hidden": 3.46640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19376826155930757,
+      "step": 1000
+    },
+    {
+      "epoch": 0.02525,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.1552039416015355e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4528,
+      "loss/crossentropy": 2.003750593960285,
+      "loss/hidden": 3.330859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18129821103066207,
+      "step": 1010
+    },
+    {
+      "epoch": 0.0255,
+      "grad_norm": 35.25,
+      "grad_norm_var": 24.095572916666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5395,
+      "loss/crossentropy": 2.0453194856643675,
+      "loss/hidden": 3.477734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.199107607267797,
+      "step": 1020
+    },
+    {
+      "epoch": 0.02575,
+      "grad_norm": 32.25,
+      "grad_norm_var": 19.5259765625,
+      "learning_rate": 0.0001,
+      "loss": 7.31,
+      "loss/crossentropy": 2.1016619503498077,
+      "loss/hidden": 3.34453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.184703135676682,
+      "step": 1030
+    },
+    {
+      "epoch": 0.026,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.87890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5425,
+      "loss/crossentropy": 2.1467826470732687,
+      "loss/hidden": 3.432421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20074132941663264,
+      "step": 1040
+    },
+    {
+      "epoch": 0.02625,
+      "grad_norm": 30.625,
+      "grad_norm_var": 0.7452473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4114,
+      "loss/crossentropy": 2.049474111199379,
+      "loss/hidden": 3.41796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20267941821366547,
+      "step": 1050
+    },
+    {
+      "epoch": 0.0265,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.124739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4845,
+      "loss/crossentropy": 2.036583887040615,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1893632340244949,
+      "step": 1060
+    },
+    {
+      "epoch": 0.02675,
+      "grad_norm": 40.75,
+      "grad_norm_var": 3.405847188209664e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3982,
+      "loss/crossentropy": 2.124411530792713,
+      "loss/hidden": 3.4484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19454579129815103,
+      "step": 1070
+    },
+    {
+      "epoch": 0.027,
+      "grad_norm": 28.25,
+      "grad_norm_var": 3.4058471885941417e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3928,
+      "loss/crossentropy": 2.0034691862761975,
+      "loss/hidden": 3.503515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21349683087319135,
+      "step": 1080
+    },
+    {
+      "epoch": 0.02725,
+      "grad_norm": 29.875,
+      "grad_norm_var": 4.88515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5095,
+      "loss/crossentropy": 1.9183670297265052,
+      "loss/hidden": 3.405859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19249978363513948,
+      "step": 1090
+    },
+    {
+      "epoch": 0.0275,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.2728515625,
+      "learning_rate": 0.0001,
+      "loss": 7.37,
+      "loss/crossentropy": 2.145428071916103,
+      "loss/hidden": 3.35703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19729665387421846,
+      "step": 1100
+    },
+    {
+      "epoch": 0.02775,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.34765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4772,
+      "loss/crossentropy": 2.10652961358428,
+      "loss/hidden": 3.398046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19585925145074726,
+      "step": 1110
+    },
+    {
+      "epoch": 0.028,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.434477049308093e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4016,
+      "loss/crossentropy": 1.9645449101924897,
+      "loss/hidden": 3.44453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19977953620254993,
+      "step": 1120
+    },
+    {
+      "epoch": 0.02825,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.4344770492950907e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4453,
+      "loss/crossentropy": 2.131172102689743,
+      "loss/hidden": 3.383984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2083016105927527,
+      "step": 1130
+    },
+    {
+      "epoch": 0.0285,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.7080729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4009,
+      "loss/crossentropy": 2.003016713261604,
+      "loss/hidden": 3.34453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18665643623098732,
+      "step": 1140
+    },
+    {
+      "epoch": 0.02875,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.34765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5648,
+      "loss/crossentropy": 2.0709651306271555,
+      "loss/hidden": 3.45703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18793081305921078,
+      "step": 1150
+    },
+    {
+      "epoch": 0.029,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.1582682291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4644,
+      "loss/crossentropy": 2.06434089243412,
+      "loss/hidden": 3.454296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2109043262898922,
+      "step": 1160
+    },
+    {
+      "epoch": 0.02925,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.4010416666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4403,
+      "loss/crossentropy": 2.0107607185840606,
+      "loss/hidden": 3.498046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20349722560495137,
+      "step": 1170
+    },
+    {
+      "epoch": 0.0295,
+      "grad_norm": 33.25,
+      "grad_norm_var": 1.2260416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4412,
+      "loss/crossentropy": 2.096436749398708,
+      "loss/hidden": 3.474609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20087064132094384,
+      "step": 1180
+    },
+    {
+      "epoch": 0.02975,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.8046223958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4458,
+      "loss/crossentropy": 1.972258360683918,
+      "loss/hidden": 3.583984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20998958311975002,
+      "step": 1190
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 33.75,
+      "grad_norm_var": 3.7395833333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3931,
+      "loss/crossentropy": 1.8556599006056786,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19810242671519518,
+      "step": 1200
+    },
+    {
+      "epoch": 0.03025,
+      "grad_norm": 29.0,
+      "grad_norm_var": 9.394791666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5849,
+      "loss/crossentropy": 2.0611833460628985,
+      "loss/hidden": 3.3984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19216072149574756,
+      "step": 1210
+    },
+    {
+      "epoch": 0.0305,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.26640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4844,
+      "loss/crossentropy": 2.0546294137835504,
+      "loss/hidden": 3.58828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21588555499911308,
+      "step": 1220
+    },
+    {
+      "epoch": 0.03075,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.3968098958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4858,
+      "loss/crossentropy": 2.0615282475948336,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.206529095210135,
+      "step": 1230
+    },
+    {
+      "epoch": 0.031,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.6124348958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4647,
+      "loss/crossentropy": 1.9786661133170127,
+      "loss/hidden": 3.381640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17899234425276517,
+      "step": 1240
+    },
+    {
+      "epoch": 0.03125,
+      "grad_norm": 5838471168.0,
+      "grad_norm_var": 2.1304840753447437e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4926,
+      "loss/crossentropy": 2.04936410933733,
+      "loss/hidden": 3.714453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1995564555749297,
+      "step": 1250
+    },
+    {
+      "epoch": 0.0315,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.1304840747304878e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5078,
+      "loss/crossentropy": 2.1189576953649523,
+      "loss/hidden": 3.43515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19967459067702292,
+      "step": 1260
+    },
+    {
+      "epoch": 0.03175,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.178580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4255,
+      "loss/crossentropy": 2.163596141338348,
+      "loss/hidden": 3.4546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19321363251656293,
+      "step": 1270
+    },
+    {
+      "epoch": 0.032,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.1639973958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4609,
+      "loss/crossentropy": 1.9938266813755035,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18334759529680014,
+      "step": 1280
+    },
+    {
+      "epoch": 0.03225,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.67890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4652,
+      "loss/crossentropy": 2.161333967000246,
+      "loss/hidden": 3.38828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19740422032773494,
+      "step": 1290
+    },
+    {
+      "epoch": 0.0325,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.0385416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3146,
+      "loss/crossentropy": 2.0165325723588468,
+      "loss/hidden": 3.49921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19117104820907116,
+      "step": 1300
+    },
+    {
+      "epoch": 0.03275,
+      "grad_norm": 28.25,
+      "grad_norm_var": 9.158072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4955,
+      "loss/crossentropy": 2.124955786764622,
+      "loss/hidden": 3.491015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19802952595055104,
+      "step": 1310
+    },
+    {
+      "epoch": 0.033,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.4535807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4311,
+      "loss/crossentropy": 2.018800371140242,
+      "loss/hidden": 3.542578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2196814114227891,
+      "step": 1320
+    },
+    {
+      "epoch": 0.03325,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.39375,
+      "learning_rate": 0.0001,
+      "loss": 7.5164,
+      "loss/crossentropy": 2.0520452961325644,
+      "loss/hidden": 3.454296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2013697015121579,
+      "step": 1330
+    },
+    {
+      "epoch": 0.0335,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.0431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5302,
+      "loss/crossentropy": 2.12932348549366,
+      "loss/hidden": 3.525,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20245677568018436,
+      "step": 1340
+    },
+    {
+      "epoch": 0.03375,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.3900390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5292,
+      "loss/crossentropy": 2.031618994474411,
+      "loss/hidden": 3.44140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19062725063413383,
+      "step": 1350
+    },
+    {
+      "epoch": 0.034,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.3447265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5755,
+      "loss/crossentropy": 2.2257011234760284,
+      "loss/hidden": 3.447265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1979327043518424,
+      "step": 1360
+    },
+    {
+      "epoch": 0.03425,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.3421223958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4219,
+      "loss/crossentropy": 2.155778780579567,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19018295016139747,
+      "step": 1370
+    },
+    {
+      "epoch": 0.0345,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.5872395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4637,
+      "loss/crossentropy": 2.058405503630638,
+      "loss/hidden": 3.39296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2114524593576789,
+      "step": 1380
+    },
+    {
+      "epoch": 0.03475,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.2994140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5834,
+      "loss/crossentropy": 2.1654782712459566,
+      "loss/hidden": 3.442578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2024593001231551,
+      "step": 1390
+    },
+    {
+      "epoch": 0.035,
+      "grad_norm": 31.125,
+      "grad_norm_var": 12.812239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4442,
+      "loss/crossentropy": 2.0921876966953277,
+      "loss/hidden": 3.286328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19270132519304753,
+      "step": 1400
+    },
+    {
+      "epoch": 0.03525,
+      "grad_norm": 29.25,
+      "grad_norm_var": 1.5108723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4779,
+      "loss/crossentropy": 1.9434148371219635,
+      "loss/hidden": 3.366015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17576389852911234,
+      "step": 1410
+    },
+    {
+      "epoch": 0.0355,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.154166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.508,
+      "loss/crossentropy": 2.0766889482736586,
+      "loss/hidden": 3.485546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20394362770020963,
+      "step": 1420
+    },
+    {
+      "epoch": 0.03575,
+      "grad_norm": 30.125,
+      "grad_norm_var": 17.580208333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4612,
+      "loss/crossentropy": 2.00380075648427,
+      "loss/hidden": 3.4453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18816210143268108,
+      "step": 1430
+    },
+    {
+      "epoch": 0.036,
+      "grad_norm": 31.375,
+      "grad_norm_var": 16.758268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4602,
+      "loss/crossentropy": 2.1938020154833793,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2016971528530121,
+      "step": 1440
+    },
+    {
+      "epoch": 0.03625,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.2556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4245,
+      "loss/crossentropy": 2.0232372283935547,
+      "loss/hidden": 3.40234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19209201391786337,
+      "step": 1450
+    },
+    {
+      "epoch": 0.0365,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.4041015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5518,
+      "loss/crossentropy": 2.2000616788864136,
+      "loss/hidden": 3.473046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22938326951116322,
+      "step": 1460
+    },
+    {
+      "epoch": 0.03675,
+      "grad_norm": 28.375,
+      "grad_norm_var": 2.0322916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4397,
+      "loss/crossentropy": 2.0838582158088683,
+      "loss/hidden": 3.451953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20685861641541123,
+      "step": 1470
+    },
+    {
+      "epoch": 0.037,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.5020833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4183,
+      "loss/crossentropy": 2.149951633810997,
+      "loss/hidden": 3.375390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1984950641170144,
+      "step": 1480
+    },
+    {
+      "epoch": 0.03725,
+      "grad_norm": 33.75,
+      "grad_norm_var": 34.10826822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.453,
+      "loss/crossentropy": 2.128306310623884,
+      "loss/hidden": 3.33203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19783397912979125,
+      "step": 1490
+    },
+    {
+      "epoch": 0.0375,
+      "grad_norm": 29.5,
+      "grad_norm_var": 5.008072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.469,
+      "loss/crossentropy": 2.042660539597273,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19274956732988358,
+      "step": 1500
+    },
+    {
+      "epoch": 0.03775,
+      "grad_norm": 33.0,
+      "grad_norm_var": 19.1775390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4119,
+      "loss/crossentropy": 2.043857058137655,
+      "loss/hidden": 3.376953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18266947232186795,
+      "step": 1510
+    },
+    {
+      "epoch": 0.038,
+      "grad_norm": 29.625,
+      "grad_norm_var": 14.303580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4362,
+      "loss/crossentropy": 1.9492302805185318,
+      "loss/hidden": 3.3515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1754497304558754,
+      "step": 1520
+    },
+    {
+      "epoch": 0.03825,
+      "grad_norm": 29.75,
+      "grad_norm_var": 23.764518229166665,
+      "learning_rate": 0.0001,
+      "loss": 7.4444,
+      "loss/crossentropy": 2.0668226674199106,
+      "loss/hidden": 3.473828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1921279976144433,
+      "step": 1530
+    },
+    {
+      "epoch": 0.0385,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.2226069790467994e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5077,
+      "loss/crossentropy": 2.1122784771025183,
+      "loss/hidden": 3.46953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22245875597000123,
+      "step": 1540
+    },
+    {
+      "epoch": 0.03875,
+      "grad_norm": 30.25,
+      "grad_norm_var": 5.382291666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4525,
+      "loss/crossentropy": 2.264697426557541,
+      "loss/hidden": 3.432421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2075907403603196,
+      "step": 1550
+    },
+    {
+      "epoch": 0.039,
+      "grad_norm": 30.0,
+      "grad_norm_var": 6.353580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5064,
+      "loss/crossentropy": 2.1150408178567885,
+      "loss/hidden": 3.5203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.23207673486322164,
+      "step": 1560
+    },
+    {
+      "epoch": 0.03925,
+      "grad_norm": 34.25,
+      "grad_norm_var": 6.72265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4578,
+      "loss/crossentropy": 2.188142140209675,
+      "loss/hidden": 3.445703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20429779235273599,
+      "step": 1570
+    },
+    {
+      "epoch": 0.0395,
+      "grad_norm": 34.75,
+      "grad_norm_var": 897.6666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.434,
+      "loss/crossentropy": 2.0795677445828913,
+      "loss/hidden": 3.3828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18706642352044583,
+      "step": 1580
+    },
+    {
+      "epoch": 0.03975,
+      "grad_norm": 28.0,
+      "grad_norm_var": 903.6327473958333,
+      "learning_rate": 0.0001,
+      "loss": 7.5655,
+      "loss/crossentropy": 2.1025844663381577,
+      "loss/hidden": 3.469140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1966788914054632,
+      "step": 1590
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 28.625,
+      "grad_norm_var": 11.97890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2578,
+      "loss/crossentropy": 2.050418493151665,
+      "loss/hidden": 3.453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20104087069630622,
+      "step": 1600
+    },
+    {
+      "epoch": 0.04025,
+      "grad_norm": 28.0,
+      "grad_norm_var": 2.255989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4393,
+      "loss/crossentropy": 2.1767756581306457,
+      "loss/hidden": 3.5140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2213939843699336,
+      "step": 1610
+    },
+    {
+      "epoch": 0.0405,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.80390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5026,
+      "loss/crossentropy": 2.126803469657898,
+      "loss/hidden": 3.39375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19106289148330688,
+      "step": 1620
+    },
+    {
+      "epoch": 0.04075,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.1249348958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4274,
+      "loss/crossentropy": 2.144256164133549,
+      "loss/hidden": 3.424609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21435861438512802,
+      "step": 1630
+    },
+    {
+      "epoch": 0.041,
+      "grad_norm": 30.25,
+      "grad_norm_var": 29.265559895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5728,
+      "loss/crossentropy": 2.2575725719332693,
+      "loss/hidden": 3.4421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20658138059079648,
+      "step": 1640
+    },
+    {
+      "epoch": 0.04125,
+      "grad_norm": 30.5,
+      "grad_norm_var": 48.35390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5776,
+      "loss/crossentropy": 2.096929042041302,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18803389491513373,
+      "step": 1650
+    },
+    {
+      "epoch": 0.0415,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.1010416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3792,
+      "loss/crossentropy": 2.0290944524109364,
+      "loss/hidden": 3.313671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19023821037262678,
+      "step": 1660
+    },
+    {
+      "epoch": 0.04175,
+      "grad_norm": 28.125,
+      "grad_norm_var": 33.49270833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5018,
+      "loss/crossentropy": 2.0678361281752586,
+      "loss/hidden": 3.35234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18862500675022603,
+      "step": 1670
+    },
+    {
+      "epoch": 0.042,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.2955729166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4432,
+      "loss/crossentropy": 2.0549797296524046,
+      "loss/hidden": 3.441796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19089050237089394,
+      "step": 1680
+    },
+    {
+      "epoch": 0.04225,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.8791666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3842,
+      "loss/crossentropy": 2.0077505365014074,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18722779098898173,
+      "step": 1690
+    },
+    {
+      "epoch": 0.0425,
+      "grad_norm": 29.375,
+      "grad_norm_var": 0.9434895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4273,
+      "loss/crossentropy": 2.071325332671404,
+      "loss/hidden": 3.486328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20270166713744403,
+      "step": 1700
+    },
+    {
+      "epoch": 0.04275,
+      "grad_norm": 38.25,
+      "grad_norm_var": 7.669791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4176,
+      "loss/crossentropy": 2.1353142291307448,
+      "loss/hidden": 3.453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19663168713450432,
+      "step": 1710
+    },
+    {
+      "epoch": 0.043,
+      "grad_norm": 28.25,
+      "grad_norm_var": 7.75,
+      "learning_rate": 0.0001,
+      "loss": 7.3818,
+      "loss/crossentropy": 1.9995346069335938,
+      "loss/hidden": 3.41015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18310597026720643,
+      "step": 1720
+    },
+    {
+      "epoch": 0.04325,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.7619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4912,
+      "loss/crossentropy": 2.1415088951587675,
+      "loss/hidden": 3.55078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22313783299177886,
+      "step": 1730
+    },
+    {
+      "epoch": 0.0435,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.0416666666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4999,
+      "loss/crossentropy": 2.1686330527067184,
+      "loss/hidden": 3.384765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20409150077030064,
+      "step": 1740
+    },
+    {
+      "epoch": 0.04375,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.724739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.438,
+      "loss/crossentropy": 1.9411263287067413,
+      "loss/hidden": 3.304296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17631518254056572,
+      "step": 1750
+    },
+    {
+      "epoch": 0.044,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.9145833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.679,
+      "loss/crossentropy": 2.1614590853452684,
+      "loss/hidden": 3.36015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.194198589771986,
+      "step": 1760
+    },
+    {
+      "epoch": 0.04425,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.039322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5095,
+      "loss/crossentropy": 2.282147654891014,
+      "loss/hidden": 3.359765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19978236705064772,
+      "step": 1770
+    },
+    {
+      "epoch": 0.0445,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.34140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5296,
+      "loss/crossentropy": 2.2078514605760575,
+      "loss/hidden": 3.403515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19668537452816964,
+      "step": 1780
+    },
+    {
+      "epoch": 0.04475,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.70390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5779,
+      "loss/crossentropy": 2.1053253799676894,
+      "loss/hidden": 3.433984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20323336366564035,
+      "step": 1790
+    },
+    {
+      "epoch": 0.045,
+      "grad_norm": 28.5,
+      "grad_norm_var": 4.8712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4866,
+      "loss/crossentropy": 2.060333488881588,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18627767637372017,
+      "step": 1800
+    },
+    {
+      "epoch": 0.04525,
+      "grad_norm": 28.0,
+      "grad_norm_var": 14.480989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5225,
+      "loss/crossentropy": 1.9755317773669958,
+      "loss/hidden": 3.54375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20334282671101392,
+      "step": 1810
+    },
+    {
+      "epoch": 0.0455,
+      "grad_norm": 29.875,
+      "grad_norm_var": 12.935872395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4781,
+      "loss/crossentropy": 2.1289859026670457,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1973018018528819,
+      "step": 1820
+    },
+    {
+      "epoch": 0.04575,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.123893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3915,
+      "loss/crossentropy": 1.9609280914068221,
+      "loss/hidden": 3.386328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1916458262130618,
+      "step": 1830
+    },
+    {
+      "epoch": 0.046,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.6332682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5095,
+      "loss/crossentropy": 2.0019985377788543,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19768325993791222,
+      "step": 1840
+    },
+    {
+      "epoch": 0.04625,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.225455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.623,
+      "loss/crossentropy": 2.0607564479112623,
+      "loss/hidden": 3.507421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20858939345926047,
+      "step": 1850
+    },
+    {
+      "epoch": 0.0465,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.9863932291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3836,
+      "loss/crossentropy": 2.132562433928251,
+      "loss/hidden": 3.40859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1956317812204361,
+      "step": 1860
+    },
+    {
+      "epoch": 0.04675,
+      "grad_norm": 36.0,
+      "grad_norm_var": 3.2171223958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4803,
+      "loss/crossentropy": 2.0316790327429772,
+      "loss/hidden": 3.396875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20630075875669718,
+      "step": 1870
+    },
+    {
+      "epoch": 0.047,
+      "grad_norm": 33.25,
+      "grad_norm_var": 16.304622395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.576,
+      "loss/crossentropy": 2.161964085698128,
+      "loss/hidden": 3.513671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21842746511101724,
+      "step": 1880
+    },
+    {
+      "epoch": 0.04725,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.3541666666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5036,
+      "loss/crossentropy": 1.8695943117141725,
+      "loss/hidden": 3.453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18793469872325658,
+      "step": 1890
+    },
+    {
+      "epoch": 0.0475,
+      "grad_norm": 34.25,
+      "grad_norm_var": 2.1780598958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5623,
+      "loss/crossentropy": 2.2376974314451217,
+      "loss/hidden": 3.489453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21696731727570295,
+      "step": 1900
+    },
+    {
+      "epoch": 0.04775,
+      "grad_norm": 30.75,
+      "grad_norm_var": 14.924934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.388,
+      "loss/crossentropy": 1.9403380863368511,
+      "loss/hidden": 3.34921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18128401823341847,
+      "step": 1910
+    },
+    {
+      "epoch": 0.048,
+      "grad_norm": 29.25,
+      "grad_norm_var": 25.1916015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4109,
+      "loss/crossentropy": 2.1744547933340073,
+      "loss/hidden": 3.423046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20097011709585785,
+      "step": 1920
+    },
+    {
+      "epoch": 0.04825,
+      "grad_norm": 29.25,
+      "grad_norm_var": 14.801822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2893,
+      "loss/crossentropy": 2.101319019496441,
+      "loss/hidden": 3.44140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1921493023633957,
+      "step": 1930
+    },
+    {
+      "epoch": 0.0485,
+      "grad_norm": 30.125,
+      "grad_norm_var": 14.517708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.579,
+      "loss/crossentropy": 2.057158224284649,
+      "loss/hidden": 3.59140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21765361074358225,
+      "step": 1940
+    },
+    {
+      "epoch": 0.04875,
+      "grad_norm": 29.625,
+      "grad_norm_var": 15.790559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3712,
+      "loss/crossentropy": 1.9415803879499436,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18346730088815094,
+      "step": 1950
+    },
+    {
+      "epoch": 0.049,
+      "grad_norm": 27.625,
+      "grad_norm_var": 9.794791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4902,
+      "loss/crossentropy": 2.035348242521286,
+      "loss/hidden": 3.439453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20268035624176264,
+      "step": 1960
+    },
+    {
+      "epoch": 0.04925,
+      "grad_norm": 35.25,
+      "grad_norm_var": 12.768684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4627,
+      "loss/crossentropy": 2.054542076587677,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2003987120464444,
+      "step": 1970
+    },
+    {
+      "epoch": 0.0495,
+      "grad_norm": 36.0,
+      "grad_norm_var": 12.572916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.353,
+      "loss/crossentropy": 1.9634785205125809,
+      "loss/hidden": 3.301171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17985089337453247,
+      "step": 1980
+    },
+    {
+      "epoch": 0.04975,
+      "grad_norm": 36.25,
+      "grad_norm_var": 9.2353515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4473,
+      "loss/crossentropy": 2.059533824026585,
+      "loss/hidden": 3.3578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19096513148397207,
+      "step": 1990
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 29.125,
+      "grad_norm_var": 13.320572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3914,
+      "loss/crossentropy": 2.011685383319855,
+      "loss/hidden": 3.4421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19188414234668016,
+      "step": 2000
+    },
+    {
+      "epoch": 0.05025,
+      "grad_norm": 36.25,
+      "grad_norm_var": 14.026822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4213,
+      "loss/crossentropy": 2.309766414761543,
+      "loss/hidden": 3.39453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20372038893401623,
+      "step": 2010
+    },
+    {
+      "epoch": 0.0505,
+      "grad_norm": 29.0,
+      "grad_norm_var": 9.237239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4145,
+      "loss/crossentropy": 2.1240487143397333,
+      "loss/hidden": 3.447265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20137840434908866,
+      "step": 2020
+    },
+    {
+      "epoch": 0.05075,
+      "grad_norm": 38.5,
+      "grad_norm_var": 89.21432291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3696,
+      "loss/crossentropy": 2.112667274475098,
+      "loss/hidden": 3.487109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19770587887614965,
+      "step": 2030
+    },
+    {
+      "epoch": 0.051,
+      "grad_norm": 27.75,
+      "grad_norm_var": 94.06015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2471,
+      "loss/crossentropy": 1.9955052442848682,
+      "loss/hidden": 3.30546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1880181163549423,
+      "step": 2040
+    },
+    {
+      "epoch": 0.05125,
+      "grad_norm": 35.25,
+      "grad_norm_var": 3.67265625,
+      "learning_rate": 0.0001,
+      "loss": 7.458,
+      "loss/crossentropy": 2.1320972844958304,
+      "loss/hidden": 3.384765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18908526431769132,
+      "step": 2050
+    },
+    {
+      "epoch": 0.0515,
+      "grad_norm": 38.75,
+      "grad_norm_var": 10.776822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3769,
+      "loss/crossentropy": 2.171598494052887,
+      "loss/hidden": 3.29765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18929236195981503,
+      "step": 2060
+    },
+    {
+      "epoch": 0.05175,
+      "grad_norm": 32.75,
+      "grad_norm_var": 10.53515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5279,
+      "loss/crossentropy": 2.0172302186489106,
+      "loss/hidden": 3.4203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2013201082125306,
+      "step": 2070
+    },
+    {
+      "epoch": 0.052,
+      "grad_norm": 32.0,
+      "grad_norm_var": 7.678125,
+      "learning_rate": 0.0001,
+      "loss": 7.3619,
+      "loss/crossentropy": 1.982726515084505,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17850281894207,
+      "step": 2080
+    },
+    {
+      "epoch": 0.05225,
+      "grad_norm": 29.75,
+      "grad_norm_var": 63.6681640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5109,
+      "loss/crossentropy": 2.121504098176956,
+      "loss/hidden": 3.50703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.240205854550004,
+      "step": 2090
+    },
+    {
+      "epoch": 0.0525,
+      "grad_norm": 34.5,
+      "grad_norm_var": 7.506184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4658,
+      "loss/crossentropy": 2.110687591135502,
+      "loss/hidden": 3.530078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2039638390764594,
+      "step": 2100
+    },
+    {
+      "epoch": 0.05275,
+      "grad_norm": 32.5,
+      "grad_norm_var": 19.075455729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5668,
+      "loss/crossentropy": 1.9557841390371322,
+      "loss/hidden": 3.462109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18774209143593906,
+      "step": 2110
+    },
+    {
+      "epoch": 0.053,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.85390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5735,
+      "loss/crossentropy": 2.0219520531594752,
+      "loss/hidden": 3.3796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18533632289618254,
+      "step": 2120
+    },
+    {
+      "epoch": 0.05325,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.8910807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4083,
+      "loss/crossentropy": 2.1359280541539194,
+      "loss/hidden": 3.412890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1897095028311014,
+      "step": 2130
+    },
+    {
+      "epoch": 0.0535,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.5957682291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.446,
+      "loss/crossentropy": 2.170258317142725,
+      "loss/hidden": 3.32109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1826348526403308,
+      "step": 2140
+    },
+    {
+      "epoch": 0.05375,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.785416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4014,
+      "loss/crossentropy": 2.131239393353462,
+      "loss/hidden": 3.303515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18656531646847724,
+      "step": 2150
+    },
+    {
+      "epoch": 0.054,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.8666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5478,
+      "loss/crossentropy": 2.223896725475788,
+      "loss/hidden": 3.383203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1951376979239285,
+      "step": 2160
+    },
+    {
+      "epoch": 0.05425,
+      "grad_norm": 30.375,
+      "grad_norm_var": 8.437955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5562,
+      "loss/crossentropy": 2.1203987300395966,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1970507999882102,
+      "step": 2170
+    },
+    {
+      "epoch": 0.0545,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.9488932291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5532,
+      "loss/crossentropy": 2.080265050381422,
+      "loss/hidden": 3.544140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2216239819303155,
+      "step": 2180
+    },
+    {
+      "epoch": 0.05475,
+      "grad_norm": 31.125,
+      "grad_norm_var": 8.1728515625,
+      "learning_rate": 0.0001,
+      "loss": 7.382,
+      "loss/crossentropy": 2.2114535331726075,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20577374435961246,
+      "step": 2190
+    },
+    {
+      "epoch": 0.055,
+      "grad_norm": 28.875,
+      "grad_norm_var": 14.520833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5766,
+      "loss/crossentropy": 2.1003271512687207,
+      "loss/hidden": 3.358984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18811229150742292,
+      "step": 2200
+    },
+    {
+      "epoch": 0.05525,
+      "grad_norm": 33.5,
+      "grad_norm_var": 16.099739583333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5553,
+      "loss/crossentropy": 2.1326127350330353,
+      "loss/hidden": 3.436328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22006579730659723,
+      "step": 2210
+    },
+    {
+      "epoch": 0.0555,
+      "grad_norm": 32.25,
+      "grad_norm_var": 9.305143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3766,
+      "loss/crossentropy": 2.1496046826243402,
+      "loss/hidden": 3.476171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1952402491122484,
+      "step": 2220
+    },
+    {
+      "epoch": 0.05575,
+      "grad_norm": 29.125,
+      "grad_norm_var": 6.805143229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3648,
+      "loss/crossentropy": 2.13938904479146,
+      "loss/hidden": 3.36640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19394674636423587,
+      "step": 2230
+    },
+    {
+      "epoch": 0.056,
+      "grad_norm": 27.625,
+      "grad_norm_var": 15.0712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4292,
+      "loss/crossentropy": 2.0648645758628845,
+      "loss/hidden": 3.436328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18520106598734856,
+      "step": 2240
+    },
+    {
+      "epoch": 0.05625,
+      "grad_norm": 29.25,
+      "grad_norm_var": 12.034309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4469,
+      "loss/crossentropy": 2.080448921024799,
+      "loss/hidden": 3.3109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18507405128329993,
+      "step": 2250
+    },
+    {
+      "epoch": 0.0565,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.014518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4325,
+      "loss/crossentropy": 2.0871294140815735,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20059894528239966,
+      "step": 2260
+    },
+    {
+      "epoch": 0.05675,
+      "grad_norm": 28.75,
+      "grad_norm_var": 1.8103515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4268,
+      "loss/crossentropy": 2.010594163835049,
+      "loss/hidden": 3.39453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19413960948586464,
+      "step": 2270
+    },
+    {
+      "epoch": 0.057,
+      "grad_norm": 32.5,
+      "grad_norm_var": 4.0369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4346,
+      "loss/crossentropy": 2.1129174560308455,
+      "loss/hidden": 3.416015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1961110396310687,
+      "step": 2280
+    },
+    {
+      "epoch": 0.05725,
+      "grad_norm": 39.0,
+      "grad_norm_var": 30.42265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4422,
+      "loss/crossentropy": 2.002947611361742,
+      "loss/hidden": 3.432421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2081361676566303,
+      "step": 2290
+    },
+    {
+      "epoch": 0.0575,
+      "grad_norm": 37.25,
+      "grad_norm_var": 25.699934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4312,
+      "loss/crossentropy": 2.06134437918663,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18918452728539706,
+      "step": 2300
+    },
+    {
+      "epoch": 0.05775,
+      "grad_norm": 28.875,
+      "grad_norm_var": 9.115559895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4209,
+      "loss/crossentropy": 2.041922479122877,
+      "loss/hidden": 3.403515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20907302405685185,
+      "step": 2310
+    },
+    {
+      "epoch": 0.058,
+      "grad_norm": 30.125,
+      "grad_norm_var": 22.248372395833332,
+      "learning_rate": 0.0001,
+      "loss": 7.6844,
+      "loss/crossentropy": 2.0152460247278214,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1905667196959257,
+      "step": 2320
+    },
+    {
+      "epoch": 0.05825,
+      "grad_norm": 38.25,
+      "grad_norm_var": 31.398893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4713,
+      "loss/crossentropy": 2.105386929959059,
+      "loss/hidden": 3.452734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1982942834496498,
+      "step": 2330
+    },
+    {
+      "epoch": 0.0585,
+      "grad_norm": 28.375,
+      "grad_norm_var": 54.94264322916667,
+      "learning_rate": 0.0001,
+      "loss": 7.4575,
+      "loss/crossentropy": 2.2358868844807147,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19232469592243434,
+      "step": 2340
+    },
+    {
+      "epoch": 0.05875,
+      "grad_norm": 33.5,
+      "grad_norm_var": 165.74583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2987,
+      "loss/crossentropy": 1.9657546751201154,
+      "loss/hidden": 3.3921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18062973748892547,
+      "step": 2350
+    },
+    {
+      "epoch": 0.059,
+      "grad_norm": 41.0,
+      "grad_norm_var": 15.376822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4431,
+      "loss/crossentropy": 2.191007924079895,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2068317520432174,
+      "step": 2360
+    },
+    {
+      "epoch": 0.05925,
+      "grad_norm": 30.625,
+      "grad_norm_var": 12.109375,
+      "learning_rate": 0.0001,
+      "loss": 7.3325,
+      "loss/crossentropy": 2.0140789330005644,
+      "loss/hidden": 3.4109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18166892379522323,
+      "step": 2370
+    },
+    {
+      "epoch": 0.0595,
+      "grad_norm": 31.875,
+      "grad_norm_var": 6.941666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4039,
+      "loss/crossentropy": 2.0221361994743345,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1934544663876295,
+      "step": 2380
+    },
+    {
+      "epoch": 0.05975,
+      "grad_norm": 30.125,
+      "grad_norm_var": 10.472330729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5862,
+      "loss/crossentropy": 1.9840030640363693,
+      "loss/hidden": 3.46640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19178631734102963,
+      "step": 2390
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 29.875,
+      "grad_norm_var": 14.10625,
+      "learning_rate": 0.0001,
+      "loss": 7.4826,
+      "loss/crossentropy": 2.1700179904699324,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1915024297311902,
+      "step": 2400
+    },
+    {
+      "epoch": 0.06025,
+      "grad_norm": 32.75,
+      "grad_norm_var": 7.370768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3889,
+      "loss/crossentropy": 2.091843403875828,
+      "loss/hidden": 3.358203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18695627991110086,
+      "step": 2410
+    },
+    {
+      "epoch": 0.0605,
+      "grad_norm": 29.0,
+      "grad_norm_var": 9.922330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4655,
+      "loss/crossentropy": 2.172381104528904,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20078962799161673,
+      "step": 2420
+    },
+    {
+      "epoch": 0.06075,
+      "grad_norm": 34.25,
+      "grad_norm_var": 8.637239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.519,
+      "loss/crossentropy": 1.995463601499796,
+      "loss/hidden": 3.411328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1993358489125967,
+      "step": 2430
+    },
+    {
+      "epoch": 0.061,
+      "grad_norm": 31.25,
+      "grad_norm_var": 11.9431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5169,
+      "loss/crossentropy": 2.296917426586151,
+      "loss/hidden": 3.513671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.23228074796497822,
+      "step": 2440
+    },
+    {
+      "epoch": 0.06125,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.4368798046573737e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5038,
+      "loss/crossentropy": 2.1944432735443113,
+      "loss/hidden": 3.3921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21073084995150565,
+      "step": 2450
+    },
+    {
+      "epoch": 0.0615,
+      "grad_norm": 33.5,
+      "grad_norm_var": 3.436879805205814e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4423,
+      "loss/crossentropy": 2.152103579044342,
+      "loss/hidden": 3.512109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20929353777319193,
+      "step": 2460
+    },
+    {
+      "epoch": 0.06175,
+      "grad_norm": 39.0,
+      "grad_norm_var": 2.2045823633093297e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4382,
+      "loss/crossentropy": 2.017627691477537,
+      "loss/hidden": 3.355078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19590776292607187,
+      "step": 2470
+    },
+    {
+      "epoch": 0.062,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.2045823636681523e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4072,
+      "loss/crossentropy": 2.1076912328600885,
+      "loss/hidden": 3.433203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1988623272627592,
+      "step": 2480
+    },
+    {
+      "epoch": 0.06225,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.2494140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3192,
+      "loss/crossentropy": 1.9777067750692368,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20539684109389783,
+      "step": 2490
+    },
+    {
+      "epoch": 0.0625,
+      "grad_norm": 29.125,
+      "grad_norm_var": 5.580208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3283,
+      "loss/crossentropy": 2.061080713570118,
+      "loss/hidden": 3.4953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20077812522649766,
+      "step": 2500
+    },
+    {
+      "epoch": 0.06275,
+      "grad_norm": 28.375,
+      "grad_norm_var": 5.618489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4401,
+      "loss/crossentropy": 2.2099071338772776,
+      "loss/hidden": 3.411328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2055276283994317,
+      "step": 2510
+    },
+    {
+      "epoch": 0.063,
+      "grad_norm": 28.125,
+      "grad_norm_var": 7.118684895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3509,
+      "loss/crossentropy": 1.962952435016632,
+      "loss/hidden": 3.421484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19731322024017572,
+      "step": 2520
+    },
+    {
+      "epoch": 0.06325,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.9681640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3695,
+      "loss/crossentropy": 1.9843583509325982,
+      "loss/hidden": 3.4921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2062232268974185,
+      "step": 2530
+    },
+    {
+      "epoch": 0.0635,
+      "grad_norm": 31.5,
+      "grad_norm_var": 3.7988932291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4485,
+      "loss/crossentropy": 2.1427679538726805,
+      "loss/hidden": 3.38125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2011977185495198,
+      "step": 2540
+    },
+    {
+      "epoch": 0.06375,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.5885416666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4157,
+      "loss/crossentropy": 1.9085583783686162,
+      "loss/hidden": 3.325,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17416954301297666,
+      "step": 2550
+    },
+    {
+      "epoch": 0.064,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.21015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5141,
+      "loss/crossentropy": 1.9622327491641045,
+      "loss/hidden": 3.361328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18756412472575903,
+      "step": 2560
+    },
+    {
+      "epoch": 0.06425,
+      "grad_norm": 30.0,
+      "grad_norm_var": 1.7143229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4624,
+      "loss/crossentropy": 2.192887546122074,
+      "loss/hidden": 3.4296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1984951412305236,
+      "step": 2570
+    },
+    {
+      "epoch": 0.0645,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.9143229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3947,
+      "loss/crossentropy": 2.102549520134926,
+      "loss/hidden": 3.463671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1989850653335452,
+      "step": 2580
+    },
+    {
+      "epoch": 0.06475,
+      "grad_norm": 32.25,
+      "grad_norm_var": 9.5322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5147,
+      "loss/crossentropy": 2.213281115144491,
+      "loss/hidden": 3.405859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2027151037938893,
+      "step": 2590
+    },
+    {
+      "epoch": 0.065,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.3427083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4691,
+      "loss/crossentropy": 2.1138279482722284,
+      "loss/hidden": 3.379296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20825629755854608,
+      "step": 2600
+    },
+    {
+      "epoch": 0.06525,
+      "grad_norm": 36.0,
+      "grad_norm_var": 3.3395182291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4775,
+      "loss/crossentropy": 2.107349547743797,
+      "loss/hidden": 3.404296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19337845854461194,
+      "step": 2610
+    },
+    {
+      "epoch": 0.0655,
+      "grad_norm": 29.25,
+      "grad_norm_var": 12.757291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5438,
+      "loss/crossentropy": 2.0628502368927,
+      "loss/hidden": 3.4984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20967572089284658,
+      "step": 2620
+    },
+    {
+      "epoch": 0.06575,
+      "grad_norm": 28.625,
+      "grad_norm_var": 11.805208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3354,
+      "loss/crossentropy": 2.1009589530527593,
+      "loss/hidden": 3.33828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18132725274190306,
+      "step": 2630
+    },
+    {
+      "epoch": 0.066,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.730208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4257,
+      "loss/crossentropy": 1.983342681080103,
+      "loss/hidden": 3.480078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19340286049991845,
+      "step": 2640
+    },
+    {
+      "epoch": 0.06625,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.7549465282226944e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.309,
+      "loss/crossentropy": 2.0057250812649725,
+      "loss/hidden": 3.418359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18936716187745334,
+      "step": 2650
+    },
+    {
+      "epoch": 0.0665,
+      "grad_norm": 36.25,
+      "grad_norm_var": 8.832747395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5442,
+      "loss/crossentropy": 2.054753464460373,
+      "loss/hidden": 3.410546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2035602940246463,
+      "step": 2660
+    },
+    {
+      "epoch": 0.06675,
+      "grad_norm": 32.5,
+      "grad_norm_var": 4.8900390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4106,
+      "loss/crossentropy": 2.0181221179664135,
+      "loss/hidden": 3.3859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1878144398331642,
+      "step": 2670
+    },
+    {
+      "epoch": 0.067,
+      "grad_norm": 30.125,
+      "grad_norm_var": 4.280989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4597,
+      "loss/crossentropy": 2.200540581345558,
+      "loss/hidden": 3.4046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20286752395331858,
+      "step": 2680
+    },
+    {
+      "epoch": 0.06725,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.8559895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4643,
+      "loss/crossentropy": 2.0630861818790436,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20401672925800085,
+      "step": 2690
+    },
+    {
+      "epoch": 0.0675,
+      "grad_norm": 33.0,
+      "grad_norm_var": 7.073958333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4001,
+      "loss/crossentropy": 1.927167509496212,
+      "loss/hidden": 3.31328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17901942003518342,
+      "step": 2700
+    },
+    {
+      "epoch": 0.06775,
+      "grad_norm": 30.25,
+      "grad_norm_var": 8.9009765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3461,
+      "loss/crossentropy": 2.0538916781544687,
+      "loss/hidden": 3.35234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1864149821922183,
+      "step": 2710
+    },
+    {
+      "epoch": 0.068,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.218489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.526,
+      "loss/crossentropy": 2.211588367819786,
+      "loss/hidden": 3.487890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20801848396658898,
+      "step": 2720
+    },
+    {
+      "epoch": 0.06825,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.0768229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5535,
+      "loss/crossentropy": 2.268890543282032,
+      "loss/hidden": 3.39921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21352684032171965,
+      "step": 2730
+    },
+    {
+      "epoch": 0.0685,
+      "grad_norm": 33.25,
+      "grad_norm_var": 5.663997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.411,
+      "loss/crossentropy": 1.902898482978344,
+      "loss/hidden": 3.423046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1794701736420393,
+      "step": 2740
+    },
+    {
+      "epoch": 0.06875,
+      "grad_norm": 32.25,
+      "grad_norm_var": 6.167708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3718,
+      "loss/crossentropy": 1.9450767874717712,
+      "loss/hidden": 3.453515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18759301900863648,
+      "step": 2750
+    },
+    {
+      "epoch": 0.069,
+      "grad_norm": 31.125,
+      "grad_norm_var": 31.185872395833332,
+      "learning_rate": 0.0001,
+      "loss": 7.4359,
+      "loss/crossentropy": 2.0783849939703942,
+      "loss/hidden": 3.334375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18503105416893958,
+      "step": 2760
+    },
+    {
+      "epoch": 0.06925,
+      "grad_norm": 36.5,
+      "grad_norm_var": 35.412434895833336,
+      "learning_rate": 0.0001,
+      "loss": 7.5806,
+      "loss/crossentropy": 2.2374701410532,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19615829903632404,
+      "step": 2770
+    },
+    {
+      "epoch": 0.0695,
+      "grad_norm": 30.25,
+      "grad_norm_var": 19.787239583333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3197,
+      "loss/crossentropy": 1.8297001466155052,
+      "loss/hidden": 3.3171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16481583826243879,
+      "step": 2780
+    },
+    {
+      "epoch": 0.06975,
+      "grad_norm": 428.0,
+      "grad_norm_var": 9873.31640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5313,
+      "loss/crossentropy": 2.249661484360695,
+      "loss/hidden": 3.392578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2018596636131406,
+      "step": 2790
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 31.0,
+      "grad_norm_var": 9755.6625,
+      "learning_rate": 0.0001,
+      "loss": 7.3957,
+      "loss/crossentropy": 1.9368772380053998,
+      "loss/hidden": 3.48203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18386599626392125,
+      "step": 2800
+    },
+    {
+      "epoch": 0.07025,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.8317057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4372,
+      "loss/crossentropy": 1.98307463824749,
+      "loss/hidden": 3.464453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19818334747105837,
+      "step": 2810
+    },
+    {
+      "epoch": 0.0705,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.589583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5014,
+      "loss/crossentropy": 2.1463105253875256,
+      "loss/hidden": 3.5046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20105676222592592,
+      "step": 2820
+    },
+    {
+      "epoch": 0.07075,
+      "grad_norm": 60.5,
+      "grad_norm_var": 178.2556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4527,
+      "loss/crossentropy": 2.0776613369584083,
+      "loss/hidden": 3.420703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19452448841184378,
+      "step": 2830
+    },
+    {
+      "epoch": 0.071,
+      "grad_norm": 29.25,
+      "grad_norm_var": 172.31451822916668,
+      "learning_rate": 0.0001,
+      "loss": 7.4802,
+      "loss/crossentropy": 2.1200039610266685,
+      "loss/hidden": 3.417578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19831879772245883,
+      "step": 2840
+    },
+    {
+      "epoch": 0.07125,
+      "grad_norm": 69.0,
+      "grad_norm_var": 117.23098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.434,
+      "loss/crossentropy": 2.024143140017986,
+      "loss/hidden": 3.348828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1836528332903981,
+      "step": 2850
+    },
+    {
+      "epoch": 0.0715,
+      "grad_norm": 31.375,
+      "grad_norm_var": 92.53723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4934,
+      "loss/crossentropy": 2.2765417456626893,
+      "loss/hidden": 3.446484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20736196860671044,
+      "step": 2860
+    },
+    {
+      "epoch": 0.07175,
+      "grad_norm": 31.625,
+      "grad_norm_var": 7.986393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4826,
+      "loss/crossentropy": 2.269197002053261,
+      "loss/hidden": 3.404296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19869209118187428,
+      "step": 2870
+    },
+    {
+      "epoch": 0.072,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.1806640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4018,
+      "loss/crossentropy": 2.2985214799642564,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20524807646870613,
+      "step": 2880
+    },
+    {
+      "epoch": 0.07225,
+      "grad_norm": 30.875,
+      "grad_norm_var": 4.801822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5148,
+      "loss/crossentropy": 2.2387808740139006,
+      "loss/hidden": 3.46015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19951685946434736,
+      "step": 2890
+    },
+    {
+      "epoch": 0.0725,
+      "grad_norm": 28.875,
+      "grad_norm_var": 13.836458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5232,
+      "loss/crossentropy": 2.049694790691137,
+      "loss/hidden": 3.410546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19052465092390775,
+      "step": 2900
+    },
+    {
+      "epoch": 0.07275,
+      "grad_norm": 29.625,
+      "grad_norm_var": 17.91640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3227,
+      "loss/crossentropy": 2.0360258772969244,
+      "loss/hidden": 3.40546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18495636582374572,
+      "step": 2910
+    },
+    {
+      "epoch": 0.073,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.8926377214767268e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4512,
+      "loss/crossentropy": 2.13848315179348,
+      "loss/hidden": 3.3859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18625867497175932,
+      "step": 2920
+    },
+    {
+      "epoch": 0.07325,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.8926377199175642e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5038,
+      "loss/crossentropy": 2.166595605015755,
+      "loss/hidden": 3.49375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20948194600641729,
+      "step": 2930
+    },
+    {
+      "epoch": 0.0735,
+      "grad_norm": 28.5,
+      "grad_norm_var": 73.08020833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.374,
+      "loss/crossentropy": 1.9849643550813199,
+      "loss/hidden": 3.301171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18302082028239966,
+      "step": 2940
+    },
+    {
+      "epoch": 0.07375,
+      "grad_norm": 29.125,
+      "grad_norm_var": 24.825,
+      "learning_rate": 0.0001,
+      "loss": 7.3651,
+      "loss/crossentropy": 2.057874396443367,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1866615541279316,
+      "step": 2950
+    },
+    {
+      "epoch": 0.074,
+      "grad_norm": 30.625,
+      "grad_norm_var": 883.6354166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5415,
+      "loss/crossentropy": 2.1631729155778885,
+      "loss/hidden": 3.388671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20762786027044058,
+      "step": 2960
+    },
+    {
+      "epoch": 0.07425,
+      "grad_norm": 32.75,
+      "grad_norm_var": 887.2705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4471,
+      "loss/crossentropy": 1.9493468508124352,
+      "loss/hidden": 3.3578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1884406829252839,
+      "step": 2970
+    },
+    {
+      "epoch": 0.0745,
+      "grad_norm": 28.875,
+      "grad_norm_var": 5.070768229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.605,
+      "loss/crossentropy": 2.122344336658716,
+      "loss/hidden": 3.460546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21057356838136912,
+      "step": 2980
+    },
+    {
+      "epoch": 0.07475,
+      "grad_norm": 37.0,
+      "grad_norm_var": 21.535416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.469,
+      "loss/crossentropy": 2.008989527821541,
+      "loss/hidden": 3.54140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2172183733433485,
+      "step": 2990
+    },
+    {
+      "epoch": 0.075,
+      "grad_norm": 29.375,
+      "grad_norm_var": 18.198958333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3932,
+      "loss/crossentropy": 2.1922819674015046,
+      "loss/hidden": 3.453515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20425879992544652,
+      "step": 3000
+    },
+    {
+      "epoch": 0.07525,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.668684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3505,
+      "loss/crossentropy": 2.189265179634094,
+      "loss/hidden": 3.34609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20808048862963915,
+      "step": 3010
+    },
+    {
+      "epoch": 0.0755,
+      "grad_norm": 30.75,
+      "grad_norm_var": 14.20625,
+      "learning_rate": 0.0001,
+      "loss": 7.5013,
+      "loss/crossentropy": 2.0573098927736284,
+      "loss/hidden": 3.3515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18116160985082388,
+      "step": 3020
+    },
+    {
+      "epoch": 0.07575,
+      "grad_norm": 31.375,
+      "grad_norm_var": 16.983333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4455,
+      "loss/crossentropy": 1.9735823571681976,
+      "loss/hidden": 3.459765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19495000168681145,
+      "step": 3030
+    },
+    {
+      "epoch": 0.076,
+      "grad_norm": 7247757312.0,
+      "grad_norm_var": 3.2831240991582193e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4881,
+      "loss/crossentropy": 1.971890377253294,
+      "loss/hidden": 3.40625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18015608433634042,
+      "step": 3040
+    },
+    {
+      "epoch": 0.07625,
+      "grad_norm": 28.25,
+      "grad_norm_var": 3.283124098780732e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3664,
+      "loss/crossentropy": 1.8378953270614147,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1741427879780531,
+      "step": 3050
+    },
+    {
+      "epoch": 0.0765,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.89140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5137,
+      "loss/crossentropy": 2.141886255145073,
+      "loss/hidden": 3.443359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19584037065505983,
+      "step": 3060
+    },
+    {
+      "epoch": 0.07675,
+      "grad_norm": 27.25,
+      "grad_norm_var": 2.4244140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4296,
+      "loss/crossentropy": 2.0373554110527037,
+      "loss/hidden": 3.5640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.216986732929945,
+      "step": 3070
+    },
+    {
+      "epoch": 0.077,
+      "grad_norm": 35.25,
+      "grad_norm_var": 3.7322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5269,
+      "loss/crossentropy": 1.975497831404209,
+      "loss/hidden": 3.333984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1780722170136869,
+      "step": 3080
+    },
+    {
+      "epoch": 0.07725,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.6895182291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4938,
+      "loss/crossentropy": 2.151789793372154,
+      "loss/hidden": 3.502734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21854450944811105,
+      "step": 3090
+    },
+    {
+      "epoch": 0.0775,
+      "grad_norm": 29.5,
+      "grad_norm_var": 6.82265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4321,
+      "loss/crossentropy": 1.9484706297516823,
+      "loss/hidden": 3.506640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19896488767117262,
+      "step": 3100
+    },
+    {
+      "epoch": 0.07775,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.0780598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5471,
+      "loss/crossentropy": 2.165594828128815,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20095103643834591,
+      "step": 3110
+    },
+    {
+      "epoch": 0.078,
+      "grad_norm": 29.0,
+      "grad_norm_var": 2.2197916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6334,
+      "loss/crossentropy": 2.1854751259088516,
+      "loss/hidden": 3.4640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21246263310313224,
+      "step": 3120
+    },
+    {
+      "epoch": 0.07825,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.71640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4278,
+      "loss/crossentropy": 1.914103902876377,
+      "loss/hidden": 3.451953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18373754434287548,
+      "step": 3130
+    },
+    {
+      "epoch": 0.0785,
+      "grad_norm": 29.0,
+      "grad_norm_var": 1.2952473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4487,
+      "loss/crossentropy": 1.9421842776238918,
+      "loss/hidden": 3.5296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19919300880283117,
+      "step": 3140
+    },
+    {
+      "epoch": 0.07875,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.8268229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5818,
+      "loss/crossentropy": 2.0765694811940194,
+      "loss/hidden": 3.5171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19946561977267266,
+      "step": 3150
+    },
+    {
+      "epoch": 0.079,
+      "grad_norm": 28.125,
+      "grad_norm_var": 11.483268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4372,
+      "loss/crossentropy": 2.013955050334334,
+      "loss/hidden": 3.4078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20109358858317136,
+      "step": 3160
+    },
+    {
+      "epoch": 0.07925,
+      "grad_norm": 28.875,
+      "grad_norm_var": 12.871809895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4606,
+      "loss/crossentropy": 2.2802242666482924,
+      "loss/hidden": 3.423046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21229397617280482,
+      "step": 3170
+    },
+    {
+      "epoch": 0.0795,
+      "grad_norm": 28.375,
+      "grad_norm_var": 1.6301432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4691,
+      "loss/crossentropy": 2.134338477253914,
+      "loss/hidden": 3.404296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18632632456719875,
+      "step": 3180
+    },
+    {
+      "epoch": 0.07975,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.6113932291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4903,
+      "loss/crossentropy": 2.192245528101921,
+      "loss/hidden": 3.381640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19276445377618073,
+      "step": 3190
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 27.875,
+      "grad_norm_var": 2.6830729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4715,
+      "loss/crossentropy": 2.1333388604223726,
+      "loss/hidden": 3.376953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1902673264965415,
+      "step": 3200
+    },
+    {
+      "epoch": 0.08025,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.7072265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4646,
+      "loss/crossentropy": 2.1069626569747926,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18933899328112602,
+      "step": 3210
+    },
+    {
+      "epoch": 0.0805,
+      "grad_norm": 33.0,
+      "grad_norm_var": 1.6457682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3771,
+      "loss/crossentropy": 2.143903985619545,
+      "loss/hidden": 3.34921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19841080270707606,
+      "step": 3220
+    },
+    {
+      "epoch": 0.08075,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.405143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4629,
+      "loss/crossentropy": 1.9501185864210129,
+      "loss/hidden": 3.474609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2003694986924529,
+      "step": 3230
+    },
+    {
+      "epoch": 0.081,
+      "grad_norm": 35.0,
+      "grad_norm_var": 3.4619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.6085,
+      "loss/crossentropy": 2.1099744185805323,
+      "loss/hidden": 3.33515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1865939747542143,
+      "step": 3240
+    },
+    {
+      "epoch": 0.08125,
+      "grad_norm": 38.0,
+      "grad_norm_var": 15.54140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4858,
+      "loss/crossentropy": 1.8915734700858593,
+      "loss/hidden": 3.550390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20414282865822314,
+      "step": 3250
+    },
+    {
+      "epoch": 0.0815,
+      "grad_norm": 31.875,
+      "grad_norm_var": 15.074934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4995,
+      "loss/crossentropy": 2.0746393710374833,
+      "loss/hidden": 3.448046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19025763403624296,
+      "step": 3260
+    },
+    {
+      "epoch": 0.08175,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.532291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4517,
+      "loss/crossentropy": 2.201898355782032,
+      "loss/hidden": 3.38515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1851862959563732,
+      "step": 3270
+    },
+    {
+      "epoch": 0.082,
+      "grad_norm": 32.25,
+      "grad_norm_var": 9.199739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4085,
+      "loss/crossentropy": 1.9774614453315735,
+      "loss/hidden": 3.5078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1853517958894372,
+      "step": 3280
+    },
+    {
+      "epoch": 0.08225,
+      "grad_norm": 31.0,
+      "grad_norm_var": 13.801497395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4065,
+      "loss/crossentropy": 2.1263367265462874,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18529028967022895,
+      "step": 3290
+    },
+    {
+      "epoch": 0.0825,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.967643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4165,
+      "loss/crossentropy": 2.193544697761536,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19897244460880756,
+      "step": 3300
+    },
+    {
+      "epoch": 0.08275,
+      "grad_norm": 33.75,
+      "grad_norm_var": 9.687239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5716,
+      "loss/crossentropy": 2.0868531957268717,
+      "loss/hidden": 3.616796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21278488002717494,
+      "step": 3310
+    },
+    {
+      "epoch": 0.083,
+      "grad_norm": 31.0,
+      "grad_norm_var": 7.9478515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5543,
+      "loss/crossentropy": 2.1392074063420297,
+      "loss/hidden": 3.395703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21113577168434858,
+      "step": 3320
+    },
+    {
+      "epoch": 0.08325,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.0268229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4454,
+      "loss/crossentropy": 2.0691144198179243,
+      "loss/hidden": 3.341796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20186964478343725,
+      "step": 3330
+    },
+    {
+      "epoch": 0.0835,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.6211653265769103e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4481,
+      "loss/crossentropy": 2.0832756504416468,
+      "loss/hidden": 3.423046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19915037509053946,
+      "step": 3340
+    },
+    {
+      "epoch": 0.08375,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.621165324337292e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3606,
+      "loss/crossentropy": 2.102260760962963,
+      "loss/hidden": 3.372265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19333885367959738,
+      "step": 3350
+    },
+    {
+      "epoch": 0.084,
+      "grad_norm": 29.25,
+      "grad_norm_var": 85.575,
+      "learning_rate": 0.0001,
+      "loss": 7.4073,
+      "loss/crossentropy": 2.149528594315052,
+      "loss/hidden": 3.491796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2177526842802763,
+      "step": 3360
+    },
+    {
+      "epoch": 0.08425,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.8645833333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4642,
+      "loss/crossentropy": 2.085590344667435,
+      "loss/hidden": 3.330859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17804578468203544,
+      "step": 3370
+    },
+    {
+      "epoch": 0.0845,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.996875,
+      "learning_rate": 0.0001,
+      "loss": 7.3953,
+      "loss/crossentropy": 2.0975965946912765,
+      "loss/hidden": 3.3,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1847201505675912,
+      "step": 3380
+    },
+    {
+      "epoch": 0.08475,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.470572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4553,
+      "loss/crossentropy": 2.1018140748143197,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18250287007540464,
+      "step": 3390
+    },
+    {
+      "epoch": 0.085,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.887955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5238,
+      "loss/crossentropy": 2.1050665065646172,
+      "loss/hidden": 3.501953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2124734738841653,
+      "step": 3400
+    },
+    {
+      "epoch": 0.08525,
+      "grad_norm": 30.0,
+      "grad_norm_var": 1.7143229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2754,
+      "loss/crossentropy": 2.0948296964168547,
+      "loss/hidden": 3.37421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.178215317055583,
+      "step": 3410
+    },
+    {
+      "epoch": 0.0855,
+      "grad_norm": 30.875,
+      "grad_norm_var": 5.354622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4191,
+      "loss/crossentropy": 2.0418393671512605,
+      "loss/hidden": 3.4734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18740264605730772,
+      "step": 3420
+    },
+    {
+      "epoch": 0.08575,
+      "grad_norm": 32.25,
+      "grad_norm_var": 6.430989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5642,
+      "loss/crossentropy": 2.0279636546969413,
+      "loss/hidden": 3.55859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20154636316001415,
+      "step": 3430
+    },
+    {
+      "epoch": 0.086,
+      "grad_norm": 29.125,
+      "grad_norm_var": 53.64791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.485,
+      "loss/crossentropy": 2.0705729112029077,
+      "loss/hidden": 3.456640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22035282999277114,
+      "step": 3440
+    },
+    {
+      "epoch": 0.08625,
+      "grad_norm": 30.375,
+      "grad_norm_var": 5.54765625,
+      "learning_rate": 0.0001,
+      "loss": 7.428,
+      "loss/crossentropy": 1.9830067940056324,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19354272997006775,
+      "step": 3450
+    },
+    {
+      "epoch": 0.0865,
+      "grad_norm": 28.375,
+      "grad_norm_var": 2.758736220726598e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4342,
+      "loss/crossentropy": 2.1590976014733316,
+      "loss/hidden": 3.489453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20231232214719058,
+      "step": 3460
+    },
+    {
+      "epoch": 0.08675,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.470572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3376,
+      "loss/crossentropy": 2.108407254517078,
+      "loss/hidden": 3.416796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18425025548785925,
+      "step": 3470
+    },
+    {
+      "epoch": 0.087,
+      "grad_norm": 32.5,
+      "grad_norm_var": 19.315559895833335,
+      "learning_rate": 0.0001,
+      "loss": 7.391,
+      "loss/crossentropy": 2.086346108466387,
+      "loss/hidden": 3.380859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19492445401847364,
+      "step": 3480
+    },
+    {
+      "epoch": 0.08725,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.9009765625,
+      "learning_rate": 0.0001,
+      "loss": 7.454,
+      "loss/crossentropy": 2.0728737086057665,
+      "loss/hidden": 3.474609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21246139723807572,
+      "step": 3490
+    },
+    {
+      "epoch": 0.0875,
+      "grad_norm": 53.25,
+      "grad_norm_var": 34.962955729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4001,
+      "loss/crossentropy": 1.9173476293683052,
+      "loss/hidden": 3.330859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18263984741643072,
+      "step": 3500
+    },
+    {
+      "epoch": 0.08775,
+      "grad_norm": 29.875,
+      "grad_norm_var": 36.22389322916667,
+      "learning_rate": 0.0001,
+      "loss": 7.5855,
+      "loss/crossentropy": 1.9761252515017986,
+      "loss/hidden": 3.391015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20959299746900797,
+      "step": 3510
+    },
+    {
+      "epoch": 0.088,
+      "grad_norm": 32.25,
+      "grad_norm_var": 17.7337890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4728,
+      "loss/crossentropy": 2.0416554152965545,
+      "loss/hidden": 3.4703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19014756735414268,
+      "step": 3520
+    },
+    {
+      "epoch": 0.08825,
+      "grad_norm": 29.375,
+      "grad_norm_var": 14.664322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5608,
+      "loss/crossentropy": 2.2834356099367143,
+      "loss/hidden": 3.3953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19908843878656626,
+      "step": 3530
+    },
+    {
+      "epoch": 0.0885,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.6702473958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4804,
+      "loss/crossentropy": 2.0417330890893934,
+      "loss/hidden": 3.46875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20852382443845272,
+      "step": 3540
+    },
+    {
+      "epoch": 0.08875,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.460724589971584e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5559,
+      "loss/crossentropy": 2.1676768481731417,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1989177169278264,
+      "step": 3550
+    },
+    {
+      "epoch": 0.089,
+      "grad_norm": 30.0,
+      "grad_norm_var": 6.881705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4678,
+      "loss/crossentropy": 2.221273897588253,
+      "loss/hidden": 3.4,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19402988757938147,
+      "step": 3560
+    },
+    {
+      "epoch": 0.08925,
+      "grad_norm": 31.375,
+      "grad_norm_var": 7.732747395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4508,
+      "loss/crossentropy": 2.1802149415016174,
+      "loss/hidden": 3.43828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20121808685362338,
+      "step": 3570
+    },
+    {
+      "epoch": 0.0895,
+      "grad_norm": 52.5,
+      "grad_norm_var": 30.9775390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3982,
+      "loss/crossentropy": 2.085124118626118,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18448642026633025,
+      "step": 3580
+    },
+    {
+      "epoch": 0.08975,
+      "grad_norm": 30.875,
+      "grad_norm_var": 32.91295572916667,
+      "learning_rate": 0.0001,
+      "loss": 7.4381,
+      "loss/crossentropy": 2.1467449337244036,
+      "loss/hidden": 3.3734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19393185302615165,
+      "step": 3590
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 29.25,
+      "grad_norm_var": 1.4708333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.415,
+      "loss/crossentropy": 2.0135369554162024,
+      "loss/hidden": 3.37109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18443848174065353,
+      "step": 3600
+    },
+    {
+      "epoch": 0.09025,
+      "grad_norm": 31.375,
+      "grad_norm_var": 6.1962890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4028,
+      "loss/crossentropy": 2.1443901300430297,
+      "loss/hidden": 3.440234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2054579086601734,
+      "step": 3610
+    },
+    {
+      "epoch": 0.0905,
+      "grad_norm": 26.5,
+      "grad_norm_var": 3.562239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3255,
+      "loss/crossentropy": 1.799356396496296,
+      "loss/hidden": 3.36953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17441922090947629,
+      "step": 3620
+    },
+    {
+      "epoch": 0.09075,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.2083333333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4272,
+      "loss/crossentropy": 1.9925116747617722,
+      "loss/hidden": 3.52578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21653544921427964,
+      "step": 3630
+    },
+    {
+      "epoch": 0.091,
+      "grad_norm": 30.125,
+      "grad_norm_var": 0.6125,
+      "learning_rate": 0.0001,
+      "loss": 7.3649,
+      "loss/crossentropy": 2.135761073231697,
+      "loss/hidden": 3.4140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18989351522177458,
+      "step": 3640
+    },
+    {
+      "epoch": 0.09125,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.4330729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4505,
+      "loss/crossentropy": 2.0986070543527604,
+      "loss/hidden": 3.334375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18352905213832854,
+      "step": 3650
+    },
+    {
+      "epoch": 0.0915,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.5869140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4199,
+      "loss/crossentropy": 2.1555575743317603,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19261632524430752,
+      "step": 3660
+    },
+    {
+      "epoch": 0.09175,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.371875,
+      "learning_rate": 0.0001,
+      "loss": 7.5463,
+      "loss/crossentropy": 2.1411691516637803,
+      "loss/hidden": 3.446875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2046652188524604,
+      "step": 3670
+    },
+    {
+      "epoch": 0.092,
+      "grad_norm": 30.625,
+      "grad_norm_var": 4.703580729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.404,
+      "loss/crossentropy": 2.142404294013977,
+      "loss/hidden": 3.445703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20414466112852098,
+      "step": 3680
+    },
+    {
+      "epoch": 0.09225,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.25625,
+      "learning_rate": 0.0001,
+      "loss": 7.4774,
+      "loss/crossentropy": 2.187901920080185,
+      "loss/hidden": 3.480859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21911972090601922,
+      "step": 3690
+    },
+    {
+      "epoch": 0.0925,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.2166666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5965,
+      "loss/crossentropy": 2.086391404271126,
+      "loss/hidden": 3.438671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2020766455680132,
+      "step": 3700
+    },
+    {
+      "epoch": 0.09275,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.147330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4579,
+      "loss/crossentropy": 2.09081457182765,
+      "loss/hidden": 3.369140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1868050311692059,
+      "step": 3710
+    },
+    {
+      "epoch": 0.093,
+      "grad_norm": 34.25,
+      "grad_norm_var": 2.467643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.522,
+      "loss/crossentropy": 2.12264247238636,
+      "loss/hidden": 3.453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18927707765251398,
+      "step": 3720
+    },
+    {
+      "epoch": 0.09325,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.981184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4155,
+      "loss/crossentropy": 2.1118928104639054,
+      "loss/hidden": 3.44765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19489197488874196,
+      "step": 3730
+    },
+    {
+      "epoch": 0.0935,
+      "grad_norm": 34.0,
+      "grad_norm_var": 5.312434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5053,
+      "loss/crossentropy": 2.1360882744193077,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19313989579677582,
+      "step": 3740
+    },
+    {
+      "epoch": 0.09375,
+      "grad_norm": 29.125,
+      "grad_norm_var": 4.549739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3275,
+      "loss/crossentropy": 2.010613538324833,
+      "loss/hidden": 3.355859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18421147018671036,
+      "step": 3750
+    },
+    {
+      "epoch": 0.094,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.5541666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4784,
+      "loss/crossentropy": 2.1465295113623144,
+      "loss/hidden": 3.323828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18987073097378016,
+      "step": 3760
+    },
+    {
+      "epoch": 0.09425,
+      "grad_norm": 32.75,
+      "grad_norm_var": 1.9018229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3495,
+      "loss/crossentropy": 2.17747982442379,
+      "loss/hidden": 3.48046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2016214355826378,
+      "step": 3770
+    },
+    {
+      "epoch": 0.0945,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.088997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5384,
+      "loss/crossentropy": 2.179350584745407,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19142594784498215,
+      "step": 3780
+    },
+    {
+      "epoch": 0.09475,
+      "grad_norm": 29.625,
+      "grad_norm_var": 1.1559895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4035,
+      "loss/crossentropy": 2.155378046631813,
+      "loss/hidden": 3.32109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19720839541405438,
+      "step": 3790
+    },
+    {
+      "epoch": 0.095,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.1999348958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4441,
+      "loss/crossentropy": 2.0597486779093743,
+      "loss/hidden": 3.413671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19279775265604257,
+      "step": 3800
+    },
+    {
+      "epoch": 0.09525,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.1666666666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5146,
+      "loss/crossentropy": 2.1966816753149034,
+      "loss/hidden": 3.44375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20174810625612735,
+      "step": 3810
+    },
+    {
+      "epoch": 0.0955,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.9593098958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.539,
+      "loss/crossentropy": 2.165803623199463,
+      "loss/hidden": 3.34375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1953417781740427,
+      "step": 3820
+    },
+    {
+      "epoch": 0.09575,
+      "grad_norm": 32.0,
+      "grad_norm_var": 6.690625,
+      "learning_rate": 0.0001,
+      "loss": 7.514,
+      "loss/crossentropy": 2.0817860513925552,
+      "loss/hidden": 3.453515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20838446952402592,
+      "step": 3830
+    },
+    {
+      "epoch": 0.096,
+      "grad_norm": 32.75,
+      "grad_norm_var": 7.6431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5472,
+      "loss/crossentropy": 2.231910442560911,
+      "loss/hidden": 3.442578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21717903479002415,
+      "step": 3840
+    },
+    {
+      "epoch": 0.09625,
+      "grad_norm": 32.0,
+      "grad_norm_var": 16.134375,
+      "learning_rate": 0.0001,
+      "loss": 7.5807,
+      "loss/crossentropy": 2.0746277555823327,
+      "loss/hidden": 3.47578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20851925816386938,
+      "step": 3850
+    },
+    {
+      "epoch": 0.0965,
+      "grad_norm": 30.625,
+      "grad_norm_var": 16.132747395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3749,
+      "loss/crossentropy": 2.1463438466191294,
+      "loss/hidden": 3.356640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19305863380432128,
+      "step": 3860
+    },
+    {
+      "epoch": 0.09675,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.0895182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5499,
+      "loss/crossentropy": 2.2108413323760034,
+      "loss/hidden": 3.42734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20310868676751853,
+      "step": 3870
+    },
+    {
+      "epoch": 0.097,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.4559895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4788,
+      "loss/crossentropy": 2.0900154620409013,
+      "loss/hidden": 3.42734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18780422061681748,
+      "step": 3880
+    },
+    {
+      "epoch": 0.09725,
+      "grad_norm": 30.625,
+      "grad_norm_var": 13.917643229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4391,
+      "loss/crossentropy": 2.0574848279356956,
+      "loss/hidden": 3.3875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19390027467161416,
+      "step": 3890
+    },
+    {
+      "epoch": 0.0975,
+      "grad_norm": 27.375,
+      "grad_norm_var": 13.55,
+      "learning_rate": 0.0001,
+      "loss": 7.4327,
+      "loss/crossentropy": 2.2832688719034193,
+      "loss/hidden": 3.3796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20608801003545524,
+      "step": 3900
+    },
+    {
+      "epoch": 0.09775,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.296875,
+      "learning_rate": 0.0001,
+      "loss": 7.3691,
+      "loss/crossentropy": 1.9183307077735663,
+      "loss/hidden": 3.37890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1917601386550814,
+      "step": 3910
+    },
+    {
+      "epoch": 0.098,
+      "grad_norm": 34.0,
+      "grad_norm_var": 3.24765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4628,
+      "loss/crossentropy": 2.0630046002566815,
+      "loss/hidden": 3.338671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1871832549571991,
+      "step": 3920
+    },
+    {
+      "epoch": 0.09825,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.5384765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4868,
+      "loss/crossentropy": 2.061261148750782,
+      "loss/hidden": 3.415234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18525551967322826,
+      "step": 3930
+    },
+    {
+      "epoch": 0.0985,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.584375,
+      "learning_rate": 0.0001,
+      "loss": 7.5498,
+      "loss/crossentropy": 2.0895790114998816,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1932330032810569,
+      "step": 3940
+    },
+    {
+      "epoch": 0.09875,
+      "grad_norm": 30.625,
+      "grad_norm_var": 25.79765625,
+      "learning_rate": 0.0001,
+      "loss": 7.6502,
+      "loss/crossentropy": 2.1616804771125318,
+      "loss/hidden": 3.365625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18905209768563508,
+      "step": 3950
+    },
+    {
+      "epoch": 0.099,
+      "grad_norm": 30.5,
+      "grad_norm_var": 28.547916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3334,
+      "loss/crossentropy": 2.1435488507151605,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1910943292081356,
+      "step": 3960
+    },
+    {
+      "epoch": 0.09925,
+      "grad_norm": 32.75,
+      "grad_norm_var": 6.3650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.542,
+      "loss/crossentropy": 2.176460310816765,
+      "loss/hidden": 3.41328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18821860365569593,
+      "step": 3970
+    },
+    {
+      "epoch": 0.0995,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.9905598958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5231,
+      "loss/crossentropy": 2.2077176332473756,
+      "loss/hidden": 3.4515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21911400128155947,
+      "step": 3980
+    },
+    {
+      "epoch": 0.09975,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.75625,
+      "learning_rate": 0.0001,
+      "loss": 7.4868,
+      "loss/crossentropy": 2.105836200714111,
+      "loss/hidden": 3.36953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1997914554551244,
+      "step": 3990
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 38.0,
+      "grad_norm_var": 4.710416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5675,
+      "loss/crossentropy": 2.233233967423439,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20876242108643056,
+      "step": 4000
+    },
+    {
+      "epoch": 0.10025,
+      "grad_norm": 28.625,
+      "grad_norm_var": 7.56640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4736,
+      "loss/crossentropy": 2.103509198874235,
+      "loss/hidden": 3.413671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1953927006572485,
+      "step": 4010
+    },
+    {
+      "epoch": 0.1005,
+      "grad_norm": 28.875,
+      "grad_norm_var": 4.119791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4509,
+      "loss/crossentropy": 1.9697775058448315,
+      "loss/hidden": 3.308984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17186311883851885,
+      "step": 4020
+    },
+    {
+      "epoch": 0.10075,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.3177083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.333,
+      "loss/crossentropy": 2.0519870311021804,
+      "loss/hidden": 3.42421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1872571600601077,
+      "step": 4030
+    },
+    {
+      "epoch": 0.101,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.2785807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3466,
+      "loss/crossentropy": 2.0663713179528713,
+      "loss/hidden": 3.39921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18582073990255593,
+      "step": 4040
+    },
+    {
+      "epoch": 0.10125,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.9577473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3812,
+      "loss/crossentropy": 2.1256399258971213,
+      "loss/hidden": 3.30390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19628962082788348,
+      "step": 4050
+    },
+    {
+      "epoch": 0.1015,
+      "grad_norm": 30.625,
+      "grad_norm_var": 0.53125,
+      "learning_rate": 0.0001,
+      "loss": 7.3726,
+      "loss/crossentropy": 2.1235328309237955,
+      "loss/hidden": 3.3703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18646292947232723,
+      "step": 4060
+    },
+    {
+      "epoch": 0.10175,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.19255952647709e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4564,
+      "loss/crossentropy": 2.0213126331567763,
+      "loss/hidden": 3.496875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19607899691909553,
+      "step": 4070
+    },
+    {
+      "epoch": 0.102,
+      "grad_norm": 28.75,
+      "grad_norm_var": 3.48515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3886,
+      "loss/crossentropy": 2.0899658009409903,
+      "loss/hidden": 3.340625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1851665174588561,
+      "step": 4080
+    },
+    {
+      "epoch": 0.10225,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.8692057291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4838,
+      "loss/crossentropy": 2.027493818849325,
+      "loss/hidden": 3.49765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19640162959694862,
+      "step": 4090
+    },
+    {
+      "epoch": 0.1025,
+      "grad_norm": 29.125,
+      "grad_norm_var": 11.762434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5099,
+      "loss/crossentropy": 2.056584618985653,
+      "loss/hidden": 3.32265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17638762388378382,
+      "step": 4100
+    },
+    {
+      "epoch": 0.10275,
+      "grad_norm": 30.125,
+      "grad_norm_var": 12.459375,
+      "learning_rate": 0.0001,
+      "loss": 7.5255,
+      "loss/crossentropy": 2.0713445380330087,
+      "loss/hidden": 3.36953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18587317056953906,
+      "step": 4110
+    },
+    {
+      "epoch": 0.103,
+      "grad_norm": 33.25,
+      "grad_norm_var": 1.9958333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4437,
+      "loss/crossentropy": 2.2338072419166566,
+      "loss/hidden": 3.36171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18814200926572083,
+      "step": 4120
+    },
+    {
+      "epoch": 0.10325,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.1259765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3184,
+      "loss/crossentropy": 2.0210259817540646,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18816483654081823,
+      "step": 4130
+    },
+    {
+      "epoch": 0.1035,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.870247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5124,
+      "loss/crossentropy": 2.0151045128703116,
+      "loss/hidden": 3.371484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19255878478288652,
+      "step": 4140
+    },
+    {
+      "epoch": 0.10375,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.3926432291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5096,
+      "loss/crossentropy": 1.9808883003890514,
+      "loss/hidden": 3.449609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19115560222417116,
+      "step": 4150
+    },
+    {
+      "epoch": 0.104,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.6979166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.549,
+      "loss/crossentropy": 2.1932784736156465,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20479805655777455,
+      "step": 4160
+    },
+    {
+      "epoch": 0.10425,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.3333333333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3875,
+      "loss/crossentropy": 1.8820222720503808,
+      "loss/hidden": 3.337109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17310038600116967,
+      "step": 4170
+    },
+    {
+      "epoch": 0.1045,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.7728515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4212,
+      "loss/crossentropy": 2.082476270198822,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19099258184432982,
+      "step": 4180
+    },
+    {
+      "epoch": 0.10475,
+      "grad_norm": 30.875,
+      "grad_norm_var": 11.408268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4991,
+      "loss/crossentropy": 2.287242355942726,
+      "loss/hidden": 3.375390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1982285875827074,
+      "step": 4190
+    },
+    {
+      "epoch": 0.105,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.999739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5959,
+      "loss/crossentropy": 2.1783332407474516,
+      "loss/hidden": 3.422265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2117959801107645,
+      "step": 4200
+    },
+    {
+      "epoch": 0.10525,
+      "grad_norm": 30.0,
+      "grad_norm_var": 4.708268229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3363,
+      "loss/crossentropy": 1.955865352600813,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18177355360239744,
+      "step": 4210
+    },
+    {
+      "epoch": 0.1055,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.0254557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4673,
+      "loss/crossentropy": 1.833389012515545,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1878132861107588,
+      "step": 4220
+    },
+    {
+      "epoch": 0.10575,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.05,
+      "learning_rate": 0.0001,
+      "loss": 7.3969,
+      "loss/crossentropy": 1.9096243590116502,
+      "loss/hidden": 3.35625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17025592969730496,
+      "step": 4230
+    },
+    {
+      "epoch": 0.106,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.82265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4638,
+      "loss/crossentropy": 2.0454175233840943,
+      "loss/hidden": 3.436328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20515710916370153,
+      "step": 4240
+    },
+    {
+      "epoch": 0.10625,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.1333333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5126,
+      "loss/crossentropy": 2.089062933623791,
+      "loss/hidden": 3.4328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19156677946448325,
+      "step": 4250
+    },
+    {
+      "epoch": 0.1065,
+      "grad_norm": 29.0,
+      "grad_norm_var": 4.311393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4468,
+      "loss/crossentropy": 2.0564094200730323,
+      "loss/hidden": 3.433984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19553639348596336,
+      "step": 4260
+    },
+    {
+      "epoch": 0.10675,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.2587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4186,
+      "loss/crossentropy": 2.13806764036417,
+      "loss/hidden": 3.3859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19822277761995793,
+      "step": 4270
+    },
+    {
+      "epoch": 0.107,
+      "grad_norm": 28.0,
+      "grad_norm_var": 1.6926432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4595,
+      "loss/crossentropy": 2.0767486467957497,
+      "loss/hidden": 3.41953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1884168043732643,
+      "step": 4280
+    },
+    {
+      "epoch": 0.10725,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.3059895833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4481,
+      "loss/crossentropy": 2.033916361629963,
+      "loss/hidden": 3.45,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20558829829096795,
+      "step": 4290
+    },
+    {
+      "epoch": 0.1075,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.9375,
+      "learning_rate": 0.0001,
+      "loss": 7.4871,
+      "loss/crossentropy": 2.078028707951307,
+      "loss/hidden": 3.37578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.188079852424562,
+      "step": 4300
+    },
+    {
+      "epoch": 0.10775,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.1020833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5379,
+      "loss/crossentropy": 2.003500834107399,
+      "loss/hidden": 3.544921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20521650360897184,
+      "step": 4310
+    },
+    {
+      "epoch": 0.108,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.8447916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3536,
+      "loss/crossentropy": 2.043112625181675,
+      "loss/hidden": 3.380859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19910661596804857,
+      "step": 4320
+    },
+    {
+      "epoch": 0.10825,
+      "grad_norm": 28.5,
+      "grad_norm_var": 4.000455729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3717,
+      "loss/crossentropy": 2.1422011658549307,
+      "loss/hidden": 3.38671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19375871792435645,
+      "step": 4330
+    },
+    {
+      "epoch": 0.1085,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.6259765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5021,
+      "loss/crossentropy": 2.131446525454521,
+      "loss/hidden": 3.480078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2063008865341544,
+      "step": 4340
+    },
+    {
+      "epoch": 0.10875,
+      "grad_norm": 32.0,
+      "grad_norm_var": 5.9525390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4749,
+      "loss/crossentropy": 2.085691845417023,
+      "loss/hidden": 3.359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1889802658930421,
+      "step": 4350
+    },
+    {
+      "epoch": 0.109,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.154166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3816,
+      "loss/crossentropy": 1.8972876839339734,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17174729090183974,
+      "step": 4360
+    },
+    {
+      "epoch": 0.10925,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.7509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4444,
+      "loss/crossentropy": 2.127763804793358,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18679574280977249,
+      "step": 4370
+    },
+    {
+      "epoch": 0.1095,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.16015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4682,
+      "loss/crossentropy": 2.1872297644615175,
+      "loss/hidden": 3.31328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.183891461789608,
+      "step": 4380
+    },
+    {
+      "epoch": 0.10975,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.3692057291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.429,
+      "loss/crossentropy": 2.19267495572567,
+      "loss/hidden": 3.394140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20111876968294382,
+      "step": 4390
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.6858723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.556,
+      "loss/crossentropy": 2.1324411287903784,
+      "loss/hidden": 3.4359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2090261412784457,
+      "step": 4400
+    },
+    {
+      "epoch": 0.11025,
+      "grad_norm": 33.5,
+      "grad_norm_var": 3.374739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4081,
+      "loss/crossentropy": 1.9800483137369156,
+      "loss/hidden": 3.584375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19881114605814218,
+      "step": 4410
+    },
+    {
+      "epoch": 0.1105,
+      "grad_norm": 31.75,
+      "grad_norm_var": 4.13515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4904,
+      "loss/crossentropy": 2.053773292154074,
+      "loss/hidden": 3.323828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18270381446927786,
+      "step": 4420
+    },
+    {
+      "epoch": 0.11075,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.3499348958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4509,
+      "loss/crossentropy": 2.0641689248383046,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19245190378278493,
+      "step": 4430
+    },
+    {
+      "epoch": 0.111,
+      "grad_norm": 33.25,
+      "grad_norm_var": 3.158333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3576,
+      "loss/crossentropy": 2.073286408931017,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1892416624352336,
+      "step": 4440
+    },
+    {
+      "epoch": 0.11125,
+      "grad_norm": 35.75,
+      "grad_norm_var": 6.167122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.456,
+      "loss/crossentropy": 2.191167525947094,
+      "loss/hidden": 3.3140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19327596500515937,
+      "step": 4450
+    },
+    {
+      "epoch": 0.1115,
+      "grad_norm": 28.0,
+      "grad_norm_var": 6.762239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4254,
+      "loss/crossentropy": 1.9917161837220192,
+      "loss/hidden": 3.48828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18673346154391765,
+      "step": 4460
+    },
+    {
+      "epoch": 0.11175,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.763541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4458,
+      "loss/crossentropy": 2.0167058646678924,
+      "loss/hidden": 3.477734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20151916183531285,
+      "step": 4470
+    },
+    {
+      "epoch": 0.112,
+      "grad_norm": 30.5,
+      "grad_norm_var": 7.175455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4057,
+      "loss/crossentropy": 2.013149876892567,
+      "loss/hidden": 3.405859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1819242848083377,
+      "step": 4480
+    },
+    {
+      "epoch": 0.11225,
+      "grad_norm": 43.25,
+      "grad_norm_var": 13.478580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4416,
+      "loss/crossentropy": 2.111778366565704,
+      "loss/hidden": 3.4,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20088088884949684,
+      "step": 4490
+    },
+    {
+      "epoch": 0.1125,
+      "grad_norm": 30.125,
+      "grad_norm_var": 11.905143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4435,
+      "loss/crossentropy": 2.0223396182060243,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1967620700597763,
+      "step": 4500
+    },
+    {
+      "epoch": 0.11275,
+      "grad_norm": 28.375,
+      "grad_norm_var": 2.2978515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3969,
+      "loss/crossentropy": 1.9966137878596784,
+      "loss/hidden": 3.416796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19062119219452142,
+      "step": 4510
+    },
+    {
+      "epoch": 0.113,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.1759765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2845,
+      "loss/crossentropy": 1.8878834903240205,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16922880560159684,
+      "step": 4520
+    },
+    {
+      "epoch": 0.11325,
+      "grad_norm": 33.5,
+      "grad_norm_var": 3.78515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5223,
+      "loss/crossentropy": 2.0424712359905244,
+      "loss/hidden": 3.42109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18261839263141155,
+      "step": 4530
+    },
+    {
+      "epoch": 0.1135,
+      "grad_norm": 41.75,
+      "grad_norm_var": 13.172330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4917,
+      "loss/crossentropy": 2.1800880253314974,
+      "loss/hidden": 3.419140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1875661849975586,
+      "step": 4540
+    },
+    {
+      "epoch": 0.11375,
+      "grad_norm": 29.5,
+      "grad_norm_var": 13.737239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4929,
+      "loss/crossentropy": 2.1130245834589005,
+      "loss/hidden": 3.514453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20742647554725407,
+      "step": 4550
+    },
+    {
+      "epoch": 0.114,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.1447265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4885,
+      "loss/crossentropy": 2.0878429099917413,
+      "loss/hidden": 3.4625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19807947240769863,
+      "step": 4560
+    },
+    {
+      "epoch": 0.11425,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.9080729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.412,
+      "loss/crossentropy": 2.045598204433918,
+      "loss/hidden": 3.43984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19935160782188177,
+      "step": 4570
+    },
+    {
+      "epoch": 0.1145,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.703285650940459e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4112,
+      "loss/crossentropy": 1.9612677067518234,
+      "loss/hidden": 3.484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1939171139150858,
+      "step": 4580
+    },
+    {
+      "epoch": 0.11475,
+      "grad_norm": 30.125,
+      "grad_norm_var": 9.067708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4109,
+      "loss/crossentropy": 2.066862888634205,
+      "loss/hidden": 3.440625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20057452656328678,
+      "step": 4590
+    },
+    {
+      "epoch": 0.115,
+      "grad_norm": 29.25,
+      "grad_norm_var": 6.670833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3857,
+      "loss/crossentropy": 2.0378803849220275,
+      "loss/hidden": 3.495703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19217969439923763,
+      "step": 4600
+    },
+    {
+      "epoch": 0.11525,
+      "grad_norm": 32.0,
+      "grad_norm_var": 8.108268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4449,
+      "loss/crossentropy": 1.9883966132998467,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1796421378850937,
+      "step": 4610
+    },
+    {
+      "epoch": 0.1155,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.8853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.43,
+      "loss/crossentropy": 2.2122382700443266,
+      "loss/hidden": 3.434765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20737907551229,
+      "step": 4620
+    },
+    {
+      "epoch": 0.11575,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.7968098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3858,
+      "loss/crossentropy": 2.0896764233708383,
+      "loss/hidden": 3.540234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20905990786850454,
+      "step": 4630
+    },
+    {
+      "epoch": 0.116,
+      "grad_norm": 27.5,
+      "grad_norm_var": 3.6879557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5145,
+      "loss/crossentropy": 2.104724445939064,
+      "loss/hidden": 3.3796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19548750538378953,
+      "step": 4640
+    },
+    {
+      "epoch": 0.11625,
+      "grad_norm": 29.875,
+      "grad_norm_var": 8.7056640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4009,
+      "loss/crossentropy": 2.155320603400469,
+      "loss/hidden": 3.47578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2002986514940858,
+      "step": 4650
+    },
+    {
+      "epoch": 0.1165,
+      "grad_norm": 27.0,
+      "grad_norm_var": 5.1541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3193,
+      "loss/crossentropy": 2.085461828112602,
+      "loss/hidden": 3.38671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1905359473079443,
+      "step": 4660
+    },
+    {
+      "epoch": 0.11675,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.5926432291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3125,
+      "loss/crossentropy": 1.9927285239100456,
+      "loss/hidden": 3.411328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17640038076788186,
+      "step": 4670
+    },
+    {
+      "epoch": 0.117,
+      "grad_norm": 33.75,
+      "grad_norm_var": 4.747330729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.469,
+      "loss/crossentropy": 2.1633560836315153,
+      "loss/hidden": 3.324609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1862495567649603,
+      "step": 4680
+    },
+    {
+      "epoch": 0.11725,
+      "grad_norm": 28.25,
+      "grad_norm_var": 7.198372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4318,
+      "loss/crossentropy": 2.2390024289488792,
+      "loss/hidden": 3.430078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2097862558439374,
+      "step": 4690
+    },
+    {
+      "epoch": 0.1175,
+      "grad_norm": 31.375,
+      "grad_norm_var": 5.760872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4669,
+      "loss/crossentropy": 2.0608770951628683,
+      "loss/hidden": 3.4390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19615320730954408,
+      "step": 4700
+    },
+    {
+      "epoch": 0.11775,
+      "grad_norm": 34.25,
+      "grad_norm_var": 4.1894735190686346e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4596,
+      "loss/crossentropy": 2.0900899082422257,
+      "loss/hidden": 3.360546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17933723451569678,
+      "step": 4710
+    },
+    {
+      "epoch": 0.118,
+      "grad_norm": 29.625,
+      "grad_norm_var": 58.10729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3979,
+      "loss/crossentropy": 2.094898019731045,
+      "loss/hidden": 3.46875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20720194689929486,
+      "step": 4720
+    },
+    {
+      "epoch": 0.11825,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.98515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4519,
+      "loss/crossentropy": 2.083225329220295,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20777787994593383,
+      "step": 4730
+    },
+    {
+      "epoch": 0.1185,
+      "grad_norm": 30.375,
+      "grad_norm_var": 4.818684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4795,
+      "loss/crossentropy": 2.1974314540624618,
+      "loss/hidden": 3.38046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19978385213762523,
+      "step": 4740
+    },
+    {
+      "epoch": 0.11875,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.439322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3843,
+      "loss/crossentropy": 1.9562335655093193,
+      "loss/hidden": 3.39140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18924889974296094,
+      "step": 4750
+    },
+    {
+      "epoch": 0.119,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.3015402743274143e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5729,
+      "loss/crossentropy": 2.0693807609379293,
+      "loss/hidden": 3.339453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18801879994571208,
+      "step": 4760
+    },
+    {
+      "epoch": 0.11925,
+      "grad_norm": 35.25,
+      "grad_norm_var": 258.8791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3013,
+      "loss/crossentropy": 2.0631250627338886,
+      "loss/hidden": 3.3703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18974527437239885,
+      "step": 4770
+    },
+    {
+      "epoch": 0.1195,
+      "grad_norm": 28.625,
+      "grad_norm_var": 301.52233072916664,
+      "learning_rate": 0.0001,
+      "loss": 7.4639,
+      "loss/crossentropy": 2.1473939388990404,
+      "loss/hidden": 3.3984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19722200892865657,
+      "step": 4780
+    },
+    {
+      "epoch": 0.11975,
+      "grad_norm": 31.125,
+      "grad_norm_var": 25.472330729166668,
+      "learning_rate": 0.0001,
+      "loss": 7.3161,
+      "loss/crossentropy": 2.1767601929605007,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20041130091995002,
+      "step": 4790
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.8580729166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3077,
+      "loss/crossentropy": 2.0214909121394156,
+      "loss/hidden": 3.38828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19553480856120586,
+      "step": 4800
+    },
+    {
+      "epoch": 0.12025,
+      "grad_norm": 34.25,
+      "grad_norm_var": 2.3666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4537,
+      "loss/crossentropy": 2.092876334488392,
+      "loss/hidden": 3.276171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19079044535756112,
+      "step": 4810
+    },
+    {
+      "epoch": 0.1205,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.1494140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3579,
+      "loss/crossentropy": 2.159788618981838,
+      "loss/hidden": 3.447265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20938555523753166,
+      "step": 4820
+    },
+    {
+      "epoch": 0.12075,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.2635411529466906e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3822,
+      "loss/crossentropy": 2.221826246380806,
+      "loss/hidden": 3.3140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18899439387023448,
+      "step": 4830
+    },
+    {
+      "epoch": 0.121,
+      "grad_norm": 29.375,
+      "grad_norm_var": 7.171875,
+      "learning_rate": 0.0001,
+      "loss": 7.3649,
+      "loss/crossentropy": 2.2076950490474703,
+      "loss/hidden": 3.321875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1911212421953678,
+      "step": 4840
+    },
+    {
+      "epoch": 0.12125,
+      "grad_norm": 28.875,
+      "grad_norm_var": 5.397916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2934,
+      "loss/crossentropy": 2.1398009806871414,
+      "loss/hidden": 3.276953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18104367554187775,
+      "step": 4850
+    },
+    {
+      "epoch": 0.1215,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.292122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3944,
+      "loss/crossentropy": 2.0568679124116898,
+      "loss/hidden": 3.31953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19066975675523282,
+      "step": 4860
+    },
+    {
+      "epoch": 0.12175,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.5145182291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5365,
+      "loss/crossentropy": 2.2600763499736787,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20988074019551278,
+      "step": 4870
+    },
+    {
+      "epoch": 0.122,
+      "grad_norm": 30.125,
+      "grad_norm_var": 0.8442057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4425,
+      "loss/crossentropy": 2.087808459997177,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20126468148082494,
+      "step": 4880
+    },
+    {
+      "epoch": 0.12225,
+      "grad_norm": 29.25,
+      "grad_norm_var": 1.9455729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4649,
+      "loss/crossentropy": 2.089573635160923,
+      "loss/hidden": 3.3890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18984669484198094,
+      "step": 4890
+    },
+    {
+      "epoch": 0.1225,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.7552083333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4894,
+      "loss/crossentropy": 2.1424145482480528,
+      "loss/hidden": 3.47890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20886036530137062,
+      "step": 4900
+    },
+    {
+      "epoch": 0.12275,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.751497395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5033,
+      "loss/crossentropy": 2.104494086652994,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1945918256416917,
+      "step": 4910
+    },
+    {
+      "epoch": 0.123,
+      "grad_norm": 28.125,
+      "grad_norm_var": 5.330989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4954,
+      "loss/crossentropy": 2.0843611776828768,
+      "loss/hidden": 3.358203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1925347488373518,
+      "step": 4920
+    },
+    {
+      "epoch": 0.12325,
+      "grad_norm": 28.625,
+      "grad_norm_var": 3.8166015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4404,
+      "loss/crossentropy": 2.205425333976746,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18580489940941333,
+      "step": 4930
+    },
+    {
+      "epoch": 0.1235,
+      "grad_norm": 29.375,
+      "grad_norm_var": 14.980208333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3481,
+      "loss/crossentropy": 1.9896500617265702,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1904701752588153,
+      "step": 4940
+    },
+    {
+      "epoch": 0.12375,
+      "grad_norm": 32.75,
+      "grad_norm_var": 19.178580729166665,
+      "learning_rate": 0.0001,
+      "loss": 7.5252,
+      "loss/crossentropy": 2.1207278318703175,
+      "loss/hidden": 3.484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19760717861354352,
+      "step": 4950
+    },
+    {
+      "epoch": 0.124,
+      "grad_norm": 32.5,
+      "grad_norm_var": 17.264583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2678,
+      "loss/crossentropy": 1.9271991185843944,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19860625620931388,
+      "step": 4960
+    },
+    {
+      "epoch": 0.12425,
+      "grad_norm": 28.625,
+      "grad_norm_var": 11.196809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3703,
+      "loss/crossentropy": 2.0659097760915754,
+      "loss/hidden": 3.287109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18224728610366583,
+      "step": 4970
+    },
+    {
+      "epoch": 0.1245,
+      "grad_norm": 37.75,
+      "grad_norm_var": 10.03515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5041,
+      "loss/crossentropy": 1.9809176340699195,
+      "loss/hidden": 3.41796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19965030066668987,
+      "step": 4980
+    },
+    {
+      "epoch": 0.12475,
+      "grad_norm": 27.125,
+      "grad_norm_var": 11.567708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.327,
+      "loss/crossentropy": 2.0197409205138683,
+      "loss/hidden": 3.368359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18525638189166785,
+      "step": 4990
+    },
+    {
+      "epoch": 0.125,
+      "grad_norm": 34.75,
+      "grad_norm_var": 8.558268229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.393,
+      "loss/crossentropy": 2.100055608153343,
+      "loss/hidden": 3.391015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19607669236138464,
+      "step": 5000
+    },
+    {
+      "epoch": 0.12525,
+      "grad_norm": 29.75,
+      "grad_norm_var": 5.602083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4165,
+      "loss/crossentropy": 1.9898378394544125,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18131834492087365,
+      "step": 5010
+    },
+    {
+      "epoch": 0.1255,
+      "grad_norm": 34.75,
+      "grad_norm_var": 5.866666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.555,
+      "loss/crossentropy": 2.086017055809498,
+      "loss/hidden": 3.446484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1889553153887391,
+      "step": 5020
+    },
+    {
+      "epoch": 0.12575,
+      "grad_norm": 33.25,
+      "grad_norm_var": 9.083268229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4098,
+      "loss/crossentropy": 2.1133529357612133,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1881294794380665,
+      "step": 5030
+    },
+    {
+      "epoch": 0.126,
+      "grad_norm": 37.25,
+      "grad_norm_var": 14.268489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4117,
+      "loss/crossentropy": 2.1818468660116195,
+      "loss/hidden": 3.3390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19351670220494271,
+      "step": 5040
+    },
+    {
+      "epoch": 0.12625,
+      "grad_norm": 27.75,
+      "grad_norm_var": 15.270572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.448,
+      "loss/crossentropy": 2.133790023624897,
+      "loss/hidden": 3.267578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17900315206497908,
+      "step": 5050
+    },
+    {
+      "epoch": 0.1265,
+      "grad_norm": 27.75,
+      "grad_norm_var": 12.469791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4514,
+      "loss/crossentropy": 2.013399636745453,
+      "loss/hidden": 3.502734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19984339475631713,
+      "step": 5060
+    },
+    {
+      "epoch": 0.12675,
+      "grad_norm": 28.625,
+      "grad_norm_var": 6.479622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4411,
+      "loss/crossentropy": 2.1552533119916917,
+      "loss/hidden": 3.4265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20999168753623962,
+      "step": 5070
+    },
+    {
+      "epoch": 0.127,
+      "grad_norm": 30.125,
+      "grad_norm_var": 4.280989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4238,
+      "loss/crossentropy": 2.1047082796692846,
+      "loss/hidden": 3.44375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19757428932935,
+      "step": 5080
+    },
+    {
+      "epoch": 0.12725,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.971875,
+      "learning_rate": 0.0001,
+      "loss": 7.547,
+      "loss/crossentropy": 2.2064288735389708,
+      "loss/hidden": 3.472265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2037733059376478,
+      "step": 5090
+    },
+    {
+      "epoch": 0.1275,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.809309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5359,
+      "loss/crossentropy": 2.307460626959801,
+      "loss/hidden": 3.405078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2208320491015911,
+      "step": 5100
+    },
+    {
+      "epoch": 0.12775,
+      "grad_norm": 30.0,
+      "grad_norm_var": 6.887434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3695,
+      "loss/crossentropy": 2.1241589702665804,
+      "loss/hidden": 3.36640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18603499811142682,
+      "step": 5110
+    },
+    {
+      "epoch": 0.128,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.8353515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4045,
+      "loss/crossentropy": 2.1248120576143266,
+      "loss/hidden": 3.364453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19218573588877916,
+      "step": 5120
+    },
+    {
+      "epoch": 0.12825,
+      "grad_norm": 30.875,
+      "grad_norm_var": 18.843489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3683,
+      "loss/crossentropy": 2.0221078641712666,
+      "loss/hidden": 3.396484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19441262539476156,
+      "step": 5130
+    },
+    {
+      "epoch": 0.1285,
+      "grad_norm": 30.25,
+      "grad_norm_var": 19.755989583333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4467,
+      "loss/crossentropy": 2.0746863678097727,
+      "loss/hidden": 3.41015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20940047055482863,
+      "step": 5140
+    },
+    {
+      "epoch": 0.12875,
+      "grad_norm": 29.75,
+      "grad_norm_var": 7.226497395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4125,
+      "loss/crossentropy": 2.127023458480835,
+      "loss/hidden": 3.3984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19320496991276742,
+      "step": 5150
+    },
+    {
+      "epoch": 0.129,
+      "grad_norm": 30.875,
+      "grad_norm_var": 8.332747395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3609,
+      "loss/crossentropy": 2.0404578357934953,
+      "loss/hidden": 3.271484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18524497244507074,
+      "step": 5160
+    },
+    {
+      "epoch": 0.12925,
+      "grad_norm": 30.375,
+      "grad_norm_var": 5.566080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3806,
+      "loss/crossentropy": 2.05174797475338,
+      "loss/hidden": 3.4125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20412184661254287,
+      "step": 5170
+    },
+    {
+      "epoch": 0.1295,
+      "grad_norm": 30.875,
+      "grad_norm_var": 72.65201822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.4596,
+      "loss/crossentropy": 2.0945761643350123,
+      "loss/hidden": 3.388671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1920377543196082,
+      "step": 5180
+    },
+    {
+      "epoch": 0.12975,
+      "grad_norm": 34.25,
+      "grad_norm_var": 1.8330729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3953,
+      "loss/crossentropy": 2.0848742216825484,
+      "loss/hidden": 3.548046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22283064387738705,
+      "step": 5190
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.4244140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5061,
+      "loss/crossentropy": 1.997230054438114,
+      "loss/hidden": 3.488671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.198976163379848,
+      "step": 5200
+    },
+    {
+      "epoch": 0.13025,
+      "grad_norm": 42.0,
+      "grad_norm_var": 9.762239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.523,
+      "loss/crossentropy": 2.169138702750206,
+      "loss/hidden": 3.458984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2180183682590723,
+      "step": 5210
+    },
+    {
+      "epoch": 0.1305,
+      "grad_norm": 27.75,
+      "grad_norm_var": 13.330989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4174,
+      "loss/crossentropy": 2.0436717979609966,
+      "loss/hidden": 3.4265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18874377477914095,
+      "step": 5220
+    },
+    {
+      "epoch": 0.13075,
+      "grad_norm": 33.5,
+      "grad_norm_var": 5.186393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4735,
+      "loss/crossentropy": 2.1061771392822264,
+      "loss/hidden": 3.3125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1841479053720832,
+      "step": 5230
+    },
+    {
+      "epoch": 0.131,
+      "grad_norm": 28.25,
+      "grad_norm_var": 3.480208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4379,
+      "loss/crossentropy": 1.9957973182201385,
+      "loss/hidden": 3.46171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19976749327033758,
+      "step": 5240
+    },
+    {
+      "epoch": 0.13125,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.2249348958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.445,
+      "loss/crossentropy": 2.089694794267416,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18123079631477595,
+      "step": 5250
+    },
+    {
+      "epoch": 0.1315,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.062239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3525,
+      "loss/crossentropy": 2.096596322953701,
+      "loss/hidden": 3.3875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17838086038827897,
+      "step": 5260
+    },
+    {
+      "epoch": 0.13175,
+      "grad_norm": 27.75,
+      "grad_norm_var": 2.6619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4042,
+      "loss/crossentropy": 2.086874121427536,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18623477015644313,
+      "step": 5270
+    },
+    {
+      "epoch": 0.132,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.8416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3634,
+      "loss/crossentropy": 2.131754931807518,
+      "loss/hidden": 3.416796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20106223467737436,
+      "step": 5280
+    },
+    {
+      "epoch": 0.13225,
+      "grad_norm": 32.5,
+      "grad_norm_var": 6.598372395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4969,
+      "loss/crossentropy": 2.1548122704029082,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20439809635281564,
+      "step": 5290
+    },
+    {
+      "epoch": 0.1325,
+      "grad_norm": 28.125,
+      "grad_norm_var": 5.252083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4373,
+      "loss/crossentropy": 2.1770398393273354,
+      "loss/hidden": 3.40078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19656166546046733,
+      "step": 5300
+    },
+    {
+      "epoch": 0.13275,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.2473307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4308,
+      "loss/crossentropy": 2.152033807337284,
+      "loss/hidden": 3.3046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18985964702442287,
+      "step": 5310
+    },
+    {
+      "epoch": 0.133,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.0041666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4327,
+      "loss/crossentropy": 2.087932828068733,
+      "loss/hidden": 3.3390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.184254783205688,
+      "step": 5320
+    },
+    {
+      "epoch": 0.13325,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.4344770491390623e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3366,
+      "loss/crossentropy": 2.034941144287586,
+      "loss/hidden": 3.38359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1904723599553108,
+      "step": 5330
+    },
+    {
+      "epoch": 0.1335,
+      "grad_norm": 31.5,
+      "grad_norm_var": 6.255989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4492,
+      "loss/crossentropy": 2.152811796963215,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18841406889259815,
+      "step": 5340
+    },
+    {
+      "epoch": 0.13375,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.6822916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.465,
+      "loss/crossentropy": 2.1877569228410723,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19606791157275438,
+      "step": 5350
+    },
+    {
+      "epoch": 0.134,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.17890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3975,
+      "loss/crossentropy": 2.0353255167603495,
+      "loss/hidden": 3.340234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1815076546743512,
+      "step": 5360
+    },
+    {
+      "epoch": 0.13425,
+      "grad_norm": 31.0,
+      "grad_norm_var": 12.4525390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3462,
+      "loss/crossentropy": 1.919140312820673,
+      "loss/hidden": 3.370703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18229803508147596,
+      "step": 5370
+    },
+    {
+      "epoch": 0.1345,
+      "grad_norm": 29.625,
+      "grad_norm_var": 7.26640625,
+      "learning_rate": 0.0001,
+      "loss": 7.506,
+      "loss/crossentropy": 2.11019846200943,
+      "loss/hidden": 3.43046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19277678560465575,
+      "step": 5380
+    },
+    {
+      "epoch": 0.13475,
+      "grad_norm": 32.0,
+      "grad_norm_var": 0.9275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3668,
+      "loss/crossentropy": 2.1108837127685547,
+      "loss/hidden": 3.493359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20013203900307416,
+      "step": 5390
+    },
+    {
+      "epoch": 0.135,
+      "grad_norm": 28.125,
+      "grad_norm_var": 1.7330729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5524,
+      "loss/crossentropy": 2.16382010653615,
+      "loss/hidden": 3.371484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1924815428443253,
+      "step": 5400
+    },
+    {
+      "epoch": 0.13525,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.5893229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5393,
+      "loss/crossentropy": 2.0622613176703455,
+      "loss/hidden": 3.498828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20035731326788664,
+      "step": 5410
+    },
+    {
+      "epoch": 0.1355,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.8499348958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3466,
+      "loss/crossentropy": 2.169532992690802,
+      "loss/hidden": 3.381640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2085475005209446,
+      "step": 5420
+    },
+    {
+      "epoch": 0.13575,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.8212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4299,
+      "loss/crossentropy": 2.056758251786232,
+      "loss/hidden": 3.460546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19217969793826342,
+      "step": 5430
+    },
+    {
+      "epoch": 0.136,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.9749348958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4334,
+      "loss/crossentropy": 2.1805212616920473,
+      "loss/hidden": 3.437109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21980819348245859,
+      "step": 5440
+    },
+    {
+      "epoch": 0.13625,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.0218098958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5382,
+      "loss/crossentropy": 2.1516773015260697,
+      "loss/hidden": 3.50703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.208776849322021,
+      "step": 5450
+    },
+    {
+      "epoch": 0.1365,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.082747395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4684,
+      "loss/crossentropy": 2.1602507561445234,
+      "loss/hidden": 3.366796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1842126866802573,
+      "step": 5460
+    },
+    {
+      "epoch": 0.13675,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.6197265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3823,
+      "loss/crossentropy": 2.081377077102661,
+      "loss/hidden": 3.38359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1906685210764408,
+      "step": 5470
+    },
+    {
+      "epoch": 0.137,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.21875,
+      "learning_rate": 0.0001,
+      "loss": 7.4128,
+      "loss/crossentropy": 2.138934540748596,
+      "loss/hidden": 3.436328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18631890565156936,
+      "step": 5480
+    },
+    {
+      "epoch": 0.13725,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.958268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.568,
+      "loss/crossentropy": 2.02208868265152,
+      "loss/hidden": 3.43828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19863407909870148,
+      "step": 5490
+    },
+    {
+      "epoch": 0.1375,
+      "grad_norm": 33.0,
+      "grad_norm_var": 6.1775390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3806,
+      "loss/crossentropy": 2.0247954726219177,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18538292730227113,
+      "step": 5500
+    },
+    {
+      "epoch": 0.13775,
+      "grad_norm": 30.625,
+      "grad_norm_var": 6.076041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6241,
+      "loss/crossentropy": 2.2269895624369385,
+      "loss/hidden": 3.445703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.197488261340186,
+      "step": 5510
+    },
+    {
+      "epoch": 0.138,
+      "grad_norm": 30.25,
+      "grad_norm_var": 4.1947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4968,
+      "loss/crossentropy": 2.077942840754986,
+      "loss/hidden": 3.420703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19135653134435415,
+      "step": 5520
+    },
+    {
+      "epoch": 0.13825,
+      "grad_norm": 27.75,
+      "grad_norm_var": 2.701041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3842,
+      "loss/crossentropy": 2.104434663057327,
+      "loss/hidden": 3.37890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18928063409402968,
+      "step": 5530
+    },
+    {
+      "epoch": 0.1385,
+      "grad_norm": 31.5,
+      "grad_norm_var": 6.4087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.443,
+      "loss/crossentropy": 2.0420807294547556,
+      "loss/hidden": 3.362890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18584198467433452,
+      "step": 5540
+    },
+    {
+      "epoch": 0.13875,
+      "grad_norm": 32.5,
+      "grad_norm_var": 4.7556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.418,
+      "loss/crossentropy": 2.0941856279969215,
+      "loss/hidden": 3.311328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18290557386353612,
+      "step": 5550
+    },
+    {
+      "epoch": 0.139,
+      "grad_norm": 34.0,
+      "grad_norm_var": 3.0947916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3594,
+      "loss/crossentropy": 2.1482032746076585,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19914243686944247,
+      "step": 5560
+    },
+    {
+      "epoch": 0.13925,
+      "grad_norm": 31.875,
+      "grad_norm_var": 4.449739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5652,
+      "loss/crossentropy": 2.150991679728031,
+      "loss/hidden": 3.4984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20609580241143705,
+      "step": 5570
+    },
+    {
+      "epoch": 0.1395,
+      "grad_norm": 30.125,
+      "grad_norm_var": 8.527083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3112,
+      "loss/crossentropy": 2.1712302803993224,
+      "loss/hidden": 3.38203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18877983894199132,
+      "step": 5580
+    },
+    {
+      "epoch": 0.13975,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.1197265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4033,
+      "loss/crossentropy": 2.1502134561538697,
+      "loss/hidden": 3.4453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21474836114794016,
+      "step": 5590
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.6809895833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5089,
+      "loss/crossentropy": 2.230164831876755,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2127472611144185,
+      "step": 5600
+    },
+    {
+      "epoch": 0.14025,
+      "grad_norm": 30.625,
+      "grad_norm_var": 52.13333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5328,
+      "loss/crossentropy": 2.1207681491971018,
+      "loss/hidden": 3.459765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20428987089544534,
+      "step": 5610
+    },
+    {
+      "epoch": 0.1405,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.037239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3972,
+      "loss/crossentropy": 2.065328547358513,
+      "loss/hidden": 3.43671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19245364069938659,
+      "step": 5620
+    },
+    {
+      "epoch": 0.14075,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.40625,
+      "learning_rate": 0.0001,
+      "loss": 7.2803,
+      "loss/crossentropy": 2.0791175961494446,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1857742078602314,
+      "step": 5630
+    },
+    {
+      "epoch": 0.141,
+      "grad_norm": 29.0,
+      "grad_norm_var": 2.687239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3324,
+      "loss/crossentropy": 2.054654690623283,
+      "loss/hidden": 3.31328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18879605047404766,
+      "step": 5640
+    },
+    {
+      "epoch": 0.14125,
+      "grad_norm": 27.625,
+      "grad_norm_var": 3.06015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4804,
+      "loss/crossentropy": 2.163857588917017,
+      "loss/hidden": 3.434375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2018537001684308,
+      "step": 5650
+    },
+    {
+      "epoch": 0.1415,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.54140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4337,
+      "loss/crossentropy": 2.1230690620839594,
+      "loss/hidden": 3.33046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18610329292714595,
+      "step": 5660
+    },
+    {
+      "epoch": 0.14175,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.0171223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.484,
+      "loss/crossentropy": 2.1232656478881835,
+      "loss/hidden": 3.306640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18209880087524652,
+      "step": 5670
+    },
+    {
+      "epoch": 0.142,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.6244140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3601,
+      "loss/crossentropy": 1.9925632011145353,
+      "loss/hidden": 3.487109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18384792990982532,
+      "step": 5680
+    },
+    {
+      "epoch": 0.14225,
+      "grad_norm": 28.875,
+      "grad_norm_var": 4.595247395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3492,
+      "loss/crossentropy": 1.9747695334255695,
+      "loss/hidden": 3.4203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18710496351122857,
+      "step": 5690
+    },
+    {
+      "epoch": 0.1425,
+      "grad_norm": 30.0,
+      "grad_norm_var": 5.160872395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.1869,
+      "loss/crossentropy": 1.9229816131293773,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1764959843829274,
+      "step": 5700
+    },
+    {
+      "epoch": 0.14275,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.1177083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5353,
+      "loss/crossentropy": 2.1759460479021073,
+      "loss/hidden": 3.442578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2011772884055972,
+      "step": 5710
+    },
+    {
+      "epoch": 0.143,
+      "grad_norm": 30.875,
+      "grad_norm_var": 19.32265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4364,
+      "loss/crossentropy": 1.9983633741736413,
+      "loss/hidden": 3.44453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2036839971318841,
+      "step": 5720
+    },
+    {
+      "epoch": 0.14325,
+      "grad_norm": 31.75,
+      "grad_norm_var": 6.330989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3473,
+      "loss/crossentropy": 2.2608693316578865,
+      "loss/hidden": 3.395703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20173839703202248,
+      "step": 5730
+    },
+    {
+      "epoch": 0.1435,
+      "grad_norm": 29.0,
+      "grad_norm_var": 30.8306640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4861,
+      "loss/crossentropy": 2.191919285058975,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1848000530153513,
+      "step": 5740
+    },
+    {
+      "epoch": 0.14375,
+      "grad_norm": 32.5,
+      "grad_norm_var": 8.167643229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5179,
+      "loss/crossentropy": 2.098912109434605,
+      "loss/hidden": 3.544921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22364525627344847,
+      "step": 5750
+    },
+    {
+      "epoch": 0.144,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.5479166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4241,
+      "loss/crossentropy": 2.089163874089718,
+      "loss/hidden": 3.39296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19125983892008663,
+      "step": 5760
+    },
+    {
+      "epoch": 0.14425,
+      "grad_norm": 29.875,
+      "grad_norm_var": 8.223958333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.45,
+      "loss/crossentropy": 2.2600366115570067,
+      "loss/hidden": 3.444140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1984367400407791,
+      "step": 5770
+    },
+    {
+      "epoch": 0.1445,
+      "grad_norm": 32.0,
+      "grad_norm_var": 14.806705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3632,
+      "loss/crossentropy": 1.9510320864617825,
+      "loss/hidden": 3.347265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18547183061018585,
+      "step": 5780
+    },
+    {
+      "epoch": 0.14475,
+      "grad_norm": 30.0,
+      "grad_norm_var": 9.655989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5503,
+      "loss/crossentropy": 2.143619356304407,
+      "loss/hidden": 3.4,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20010631643235682,
+      "step": 5790
+    },
+    {
+      "epoch": 0.145,
+      "grad_norm": 28.375,
+      "grad_norm_var": 9.556184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5076,
+      "loss/crossentropy": 1.9316529139876366,
+      "loss/hidden": 3.40546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18306834027171134,
+      "step": 5800
+    },
+    {
+      "epoch": 0.14525,
+      "grad_norm": 37.25,
+      "grad_norm_var": 10.718489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5004,
+      "loss/crossentropy": 2.136544609069824,
+      "loss/hidden": 3.3453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20682645812630654,
+      "step": 5810
+    },
+    {
+      "epoch": 0.1455,
+      "grad_norm": 34.75,
+      "grad_norm_var": 9.395572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4109,
+      "loss/crossentropy": 2.0811544865369798,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.187607554346323,
+      "step": 5820
+    },
+    {
+      "epoch": 0.14575,
+      "grad_norm": 30.125,
+      "grad_norm_var": 11.476041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4615,
+      "loss/crossentropy": 2.2464685887098312,
+      "loss/hidden": 3.419140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21834317222237587,
+      "step": 5830
+    },
+    {
+      "epoch": 0.146,
+      "grad_norm": 31.875,
+      "grad_norm_var": 8.106705729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.379,
+      "loss/crossentropy": 2.1494223892688753,
+      "loss/hidden": 3.4875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20599585752934219,
+      "step": 5840
+    },
+    {
+      "epoch": 0.14625,
+      "grad_norm": 32.25,
+      "grad_norm_var": 119.0541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4223,
+      "loss/crossentropy": 2.013238602876663,
+      "loss/hidden": 3.355859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17934355642646552,
+      "step": 5850
+    },
+    {
+      "epoch": 0.1465,
+      "grad_norm": 55.5,
+      "grad_norm_var": 40.619205729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4766,
+      "loss/crossentropy": 2.1309464499354362,
+      "loss/hidden": 3.40625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1953899236395955,
+      "step": 5860
+    },
+    {
+      "epoch": 0.14675,
+      "grad_norm": 31.125,
+      "grad_norm_var": 51.764322916666664,
+      "learning_rate": 0.0001,
+      "loss": 7.5385,
+      "loss/crossentropy": 2.203585295379162,
+      "loss/hidden": 3.40625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20035494081676006,
+      "step": 5870
+    },
+    {
+      "epoch": 0.147,
+      "grad_norm": 30.125,
+      "grad_norm_var": 8.2541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3861,
+      "loss/crossentropy": 2.057890709489584,
+      "loss/hidden": 3.3515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18680873457342387,
+      "step": 5880
+    },
+    {
+      "epoch": 0.14725,
+      "grad_norm": 30.25,
+      "grad_norm_var": 4.555989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4283,
+      "loss/crossentropy": 2.049139867722988,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1856512013822794,
+      "step": 5890
+    },
+    {
+      "epoch": 0.1475,
+      "grad_norm": 30.375,
+      "grad_norm_var": 10.570768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4651,
+      "loss/crossentropy": 2.0553019613027574,
+      "loss/hidden": 3.404296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18676785845309496,
+      "step": 5900
+    },
+    {
+      "epoch": 0.14775,
+      "grad_norm": 30.125,
+      "grad_norm_var": 14.59140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4727,
+      "loss/crossentropy": 2.0098409935832025,
+      "loss/hidden": 3.48125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20080684809945523,
+      "step": 5910
+    },
+    {
+      "epoch": 0.148,
+      "grad_norm": 28.625,
+      "grad_norm_var": 8.489322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4203,
+      "loss/crossentropy": 2.2615666806697847,
+      "loss/hidden": 3.3,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18850413355976342,
+      "step": 5920
+    },
+    {
+      "epoch": 0.14825,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.6233723958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4441,
+      "loss/crossentropy": 2.178256964683533,
+      "loss/hidden": 3.29609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19248567353934048,
+      "step": 5930
+    },
+    {
+      "epoch": 0.1485,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.455208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4573,
+      "loss/crossentropy": 2.246034747362137,
+      "loss/hidden": 3.45859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2096735591068864,
+      "step": 5940
+    },
+    {
+      "epoch": 0.14875,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.8494140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4811,
+      "loss/crossentropy": 2.180899788439274,
+      "loss/hidden": 3.341796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19460927378386259,
+      "step": 5950
+    },
+    {
+      "epoch": 0.149,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.4385416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3829,
+      "loss/crossentropy": 2.258976912498474,
+      "loss/hidden": 3.332421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19133044108748437,
+      "step": 5960
+    },
+    {
+      "epoch": 0.14925,
+      "grad_norm": 29.0,
+      "grad_norm_var": 13.80390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5384,
+      "loss/crossentropy": 2.012222741544247,
+      "loss/hidden": 3.4640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19505210760980846,
+      "step": 5970
+    },
+    {
+      "epoch": 0.1495,
+      "grad_norm": 27.875,
+      "grad_norm_var": 12.9369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4033,
+      "loss/crossentropy": 2.0392286255955696,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1930427584797144,
+      "step": 5980
+    },
+    {
+      "epoch": 0.14975,
+      "grad_norm": 28.25,
+      "grad_norm_var": 18.174739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.399,
+      "loss/crossentropy": 1.9029529005289079,
+      "loss/hidden": 3.462109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.198425155505538,
+      "step": 5990
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.0737770860662226e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4994,
+      "loss/crossentropy": 1.8985859856009484,
+      "loss/hidden": 3.4609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1951824951916933,
+      "step": 6000
+    },
+    {
+      "epoch": 0.15025,
+      "grad_norm": 36.0,
+      "grad_norm_var": 3.073777086665239e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4659,
+      "loss/crossentropy": 2.097201645374298,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18253911342471837,
+      "step": 6010
+    },
+    {
+      "epoch": 0.1505,
+      "grad_norm": 27.875,
+      "grad_norm_var": 6.801041666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2415,
+      "loss/crossentropy": 2.0210610911250115,
+      "loss/hidden": 3.484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.197306059114635,
+      "step": 6020
+    },
+    {
+      "epoch": 0.15075,
+      "grad_norm": 31.0,
+      "grad_norm_var": 14.46640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4519,
+      "loss/crossentropy": 2.1985476523637772,
+      "loss/hidden": 3.475,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20262509360909461,
+      "step": 6030
+    },
+    {
+      "epoch": 0.151,
+      "grad_norm": 30.625,
+      "grad_norm_var": 6.254622395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3353,
+      "loss/crossentropy": 2.0093181416392327,
+      "loss/hidden": 3.316015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17620250331237913,
+      "step": 6040
+    },
+    {
+      "epoch": 0.15125,
+      "grad_norm": 31.625,
+      "grad_norm_var": 56.4291015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4034,
+      "loss/crossentropy": 2.177773226797581,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20441538300365208,
+      "step": 6050
+    },
+    {
+      "epoch": 0.1515,
+      "grad_norm": 26.75,
+      "grad_norm_var": 55.889322916666664,
+      "learning_rate": 0.0001,
+      "loss": 7.3245,
+      "loss/crossentropy": 2.1666259437799456,
+      "loss/hidden": 3.392578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19311951845884323,
+      "step": 6060
+    },
+    {
+      "epoch": 0.15175,
+      "grad_norm": 30.25,
+      "grad_norm_var": 91.0103515625,
+      "learning_rate": 0.0001,
+      "loss": 7.368,
+      "loss/crossentropy": 2.063462796807289,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1834208857268095,
+      "step": 6070
+    },
+    {
+      "epoch": 0.152,
+      "grad_norm": 30.5,
+      "grad_norm_var": 18.1212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4335,
+      "loss/crossentropy": 1.9907098844647408,
+      "loss/hidden": 3.53671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20707368329167367,
+      "step": 6080
+    },
+    {
+      "epoch": 0.15225,
+      "grad_norm": 38.25,
+      "grad_norm_var": 11.470247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3789,
+      "loss/crossentropy": 2.083692157268524,
+      "loss/hidden": 3.465625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1846569798886776,
+      "step": 6090
+    },
+    {
+      "epoch": 0.1525,
+      "grad_norm": 28.25,
+      "grad_norm_var": 21.829622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3767,
+      "loss/crossentropy": 2.1113929279148578,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1990992769598961,
+      "step": 6100
+    },
+    {
+      "epoch": 0.15275,
+      "grad_norm": 33.75,
+      "grad_norm_var": 20.676497395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.339,
+      "loss/crossentropy": 2.1296695113182067,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1936045665293932,
+      "step": 6110
+    },
+    {
+      "epoch": 0.153,
+      "grad_norm": 36.25,
+      "grad_norm_var": 7.8166015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5027,
+      "loss/crossentropy": 2.1011226207017897,
+      "loss/hidden": 3.471875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20695240292698144,
+      "step": 6120
+    },
+    {
+      "epoch": 0.15325,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.1304840750224113e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.506,
+      "loss/crossentropy": 2.2427969723939896,
+      "loss/hidden": 3.40625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1966634625568986,
+      "step": 6130
+    },
+    {
+      "epoch": 0.1535,
+      "grad_norm": 34.25,
+      "grad_norm_var": 36.542643229166664,
+      "learning_rate": 0.0001,
+      "loss": 7.4413,
+      "loss/crossentropy": 2.0855264641344546,
+      "loss/hidden": 3.384765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19436217453330756,
+      "step": 6140
+    },
+    {
+      "epoch": 0.15375,
+      "grad_norm": 32.75,
+      "grad_norm_var": 10.153580729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3096,
+      "loss/crossentropy": 2.0322439685463904,
+      "loss/hidden": 3.3453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1753252800554037,
+      "step": 6150
+    },
+    {
+      "epoch": 0.154,
+      "grad_norm": 31.0,
+      "grad_norm_var": 10.216080729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2481,
+      "loss/crossentropy": 2.074477408081293,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17578690703958272,
+      "step": 6160
+    },
+    {
+      "epoch": 0.15425,
+      "grad_norm": 33.25,
+      "grad_norm_var": 28.79765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4403,
+      "loss/crossentropy": 2.0863804474473,
+      "loss/hidden": 3.431640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20699662044644357,
+      "step": 6170
+    },
+    {
+      "epoch": 0.1545,
+      "grad_norm": 33.5,
+      "grad_norm_var": 24.84375,
+      "learning_rate": 0.0001,
+      "loss": 7.4609,
+      "loss/crossentropy": 2.0696858704090118,
+      "loss/hidden": 3.458984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2128069180995226,
+      "step": 6180
+    },
+    {
+      "epoch": 0.15475,
+      "grad_norm": 31.375,
+      "grad_norm_var": 5.099739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2575,
+      "loss/crossentropy": 2.182169410586357,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.181897877715528,
+      "step": 6190
+    },
+    {
+      "epoch": 0.155,
+      "grad_norm": 34.25,
+      "grad_norm_var": 4.699934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4975,
+      "loss/crossentropy": 2.165008749067783,
+      "loss/hidden": 3.492578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19800901636481286,
+      "step": 6200
+    },
+    {
+      "epoch": 0.15525,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.7738932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3992,
+      "loss/crossentropy": 2.0653695166110992,
+      "loss/hidden": 3.406640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19180236533284187,
+      "step": 6210
+    },
+    {
+      "epoch": 0.1555,
+      "grad_norm": 35.25,
+      "grad_norm_var": 5.31015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3638,
+      "loss/crossentropy": 2.1244568385183813,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18238217020407319,
+      "step": 6220
+    },
+    {
+      "epoch": 0.15575,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.2817057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5073,
+      "loss/crossentropy": 2.1881898671388624,
+      "loss/hidden": 3.498828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2045454490929842,
+      "step": 6230
+    },
+    {
+      "epoch": 0.156,
+      "grad_norm": 28.25,
+      "grad_norm_var": 2.6572265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4609,
+      "loss/crossentropy": 2.14600064009428,
+      "loss/hidden": 3.4265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18945380430668593,
+      "step": 6240
+    },
+    {
+      "epoch": 0.15625,
+      "grad_norm": 30.75,
+      "grad_norm_var": 35.73118489583333,
+      "learning_rate": 0.0001,
+      "loss": 7.3786,
+      "loss/crossentropy": 2.168429624289274,
+      "loss/hidden": 3.296484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18439108245074748,
+      "step": 6250
+    },
+    {
+      "epoch": 0.1565,
+      "grad_norm": 52.5,
+      "grad_norm_var": 64.9962890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3511,
+      "loss/crossentropy": 2.1293379329144955,
+      "loss/hidden": 3.414453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1828605517745018,
+      "step": 6260
+    },
+    {
+      "epoch": 0.15675,
+      "grad_norm": 29.25,
+      "grad_norm_var": 59.703125,
+      "learning_rate": 0.0001,
+      "loss": 7.3978,
+      "loss/crossentropy": 1.8641120925545693,
+      "loss/hidden": 3.430078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18577109538018705,
+      "step": 6270
+    },
+    {
+      "epoch": 0.157,
+      "grad_norm": 28.75,
+      "grad_norm_var": 32.1994140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4066,
+      "loss/crossentropy": 2.0997040398418902,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18688563201576472,
+      "step": 6280
+    },
+    {
+      "epoch": 0.15725,
+      "grad_norm": 31.375,
+      "grad_norm_var": 17.302018229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4043,
+      "loss/crossentropy": 1.9712626039981842,
+      "loss/hidden": 3.45078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20054549565538765,
+      "step": 6290
+    },
+    {
+      "epoch": 0.1575,
+      "grad_norm": 31.5,
+      "grad_norm_var": 17.8431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4502,
+      "loss/crossentropy": 2.0252815186977386,
+      "loss/hidden": 3.3890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18488127905875446,
+      "step": 6300
+    },
+    {
+      "epoch": 0.15775,
+      "grad_norm": 30.75,
+      "grad_norm_var": 7.995572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3829,
+      "loss/crossentropy": 2.030302118510008,
+      "loss/hidden": 3.400390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19101340658962726,
+      "step": 6310
+    },
+    {
+      "epoch": 0.158,
+      "grad_norm": 30.125,
+      "grad_norm_var": 5.805143229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3852,
+      "loss/crossentropy": 1.9795936658978461,
+      "loss/hidden": 3.45703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19110800279304385,
+      "step": 6320
+    },
+    {
+      "epoch": 0.15825,
+      "grad_norm": 34.0,
+      "grad_norm_var": 6.91640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4417,
+      "loss/crossentropy": 2.0620448149740698,
+      "loss/hidden": 3.454296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.194018579646945,
+      "step": 6330
+    },
+    {
+      "epoch": 0.1585,
+      "grad_norm": 28.125,
+      "grad_norm_var": 31.058268229166668,
+      "learning_rate": 0.0001,
+      "loss": 7.4142,
+      "loss/crossentropy": 2.012200343608856,
+      "loss/hidden": 3.4921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21260247621685266,
+      "step": 6340
+    },
+    {
+      "epoch": 0.15875,
+      "grad_norm": 36.75,
+      "grad_norm_var": 35.18118489583333,
+      "learning_rate": 0.0001,
+      "loss": 7.3776,
+      "loss/crossentropy": 1.9757203698158263,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.200297892652452,
+      "step": 6350
+    },
+    {
+      "epoch": 0.159,
+      "grad_norm": 31.125,
+      "grad_norm_var": 17.764583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4762,
+      "loss/crossentropy": 2.195678301155567,
+      "loss/hidden": 3.434375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1989523505792022,
+      "step": 6360
+    },
+    {
+      "epoch": 0.15925,
+      "grad_norm": 29.625,
+      "grad_norm_var": 12.851041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4661,
+      "loss/crossentropy": 2.0537394002079963,
+      "loss/hidden": 3.421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20311654023826123,
+      "step": 6370
+    },
+    {
+      "epoch": 0.1595,
+      "grad_norm": 30.0,
+      "grad_norm_var": 10.0994140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2759,
+      "loss/crossentropy": 2.02838040292263,
+      "loss/hidden": 3.45859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19148585237562657,
+      "step": 6380
+    },
+    {
+      "epoch": 0.15975,
+      "grad_norm": 39.0,
+      "grad_norm_var": 2324.6707682291667,
+      "learning_rate": 0.0001,
+      "loss": 7.3973,
+      "loss/crossentropy": 2.0951177358627318,
+      "loss/hidden": 3.430859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2158473737537861,
+      "step": 6390
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 40.0,
+      "grad_norm_var": 21.121809895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2877,
+      "loss/crossentropy": 1.878954614698887,
+      "loss/hidden": 3.390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18514612764120103,
+      "step": 6400
+    },
+    {
+      "epoch": 0.16025,
+      "grad_norm": 33.5,
+      "grad_norm_var": 23.1666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3598,
+      "loss/crossentropy": 2.123918867111206,
+      "loss/hidden": 3.3140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18728599287569522,
+      "step": 6410
+    },
+    {
+      "epoch": 0.1605,
+      "grad_norm": 29.25,
+      "grad_norm_var": 11.230143229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.476,
+      "loss/crossentropy": 2.168968527019024,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20153266489505767,
+      "step": 6420
+    },
+    {
+      "epoch": 0.16075,
+      "grad_norm": 30.0,
+      "grad_norm_var": 90.8056640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4289,
+      "loss/crossentropy": 2.0426762118935584,
+      "loss/hidden": 3.367578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19033107869327068,
+      "step": 6430
+    },
+    {
+      "epoch": 0.161,
+      "grad_norm": 38.25,
+      "grad_norm_var": 15.570247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3992,
+      "loss/crossentropy": 2.0535445332527162,
+      "loss/hidden": 3.453515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19420330366119742,
+      "step": 6440
+    },
+    {
+      "epoch": 0.16125,
+      "grad_norm": 45.5,
+      "grad_norm_var": 30.326822916666668,
+      "learning_rate": 0.0001,
+      "loss": 7.3881,
+      "loss/crossentropy": 1.949498599767685,
+      "loss/hidden": 3.337890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18119702748954297,
+      "step": 6450
+    },
+    {
+      "epoch": 0.1615,
+      "grad_norm": 30.625,
+      "grad_norm_var": 96.64837239583333,
+      "learning_rate": 0.0001,
+      "loss": 7.432,
+      "loss/crossentropy": 2.2534308552742006,
+      "loss/hidden": 3.334375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20010371711105107,
+      "step": 6460
+    },
+    {
+      "epoch": 0.16175,
+      "grad_norm": 34.0,
+      "grad_norm_var": 82.66015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3555,
+      "loss/crossentropy": 2.114310759305954,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19822147954255342,
+      "step": 6470
+    },
+    {
+      "epoch": 0.162,
+      "grad_norm": 29.875,
+      "grad_norm_var": 8.4541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.1893,
+      "loss/crossentropy": 2.062894639372826,
+      "loss/hidden": 3.348828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17771479729562997,
+      "step": 6480
+    },
+    {
+      "epoch": 0.16225,
+      "grad_norm": 30.375,
+      "grad_norm_var": 15.72265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3665,
+      "loss/crossentropy": 2.0109994761645793,
+      "loss/hidden": 3.45625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19815738410688938,
+      "step": 6490
+    },
+    {
+      "epoch": 0.1625,
+      "grad_norm": 33.75,
+      "grad_norm_var": 162.74680989583334,
+      "learning_rate": 0.0001,
+      "loss": 7.4105,
+      "loss/crossentropy": 2.100720961391926,
+      "loss/hidden": 3.26484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18080853056162596,
+      "step": 6500
+    },
+    {
+      "epoch": 0.16275,
+      "grad_norm": 36.0,
+      "grad_norm_var": 10.530143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4993,
+      "loss/crossentropy": 2.208073277771473,
+      "loss/hidden": 3.3328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18481182418763636,
+      "step": 6510
+    },
+    {
+      "epoch": 0.163,
+      "grad_norm": 37.0,
+      "grad_norm_var": 8.981705729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5196,
+      "loss/crossentropy": 2.2666310742497444,
+      "loss/hidden": 3.40703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20527655016630889,
+      "step": 6520
+    },
+    {
+      "epoch": 0.16325,
+      "grad_norm": 29.625,
+      "grad_norm_var": 13.959830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4239,
+      "loss/crossentropy": 2.2184250116348267,
+      "loss/hidden": 3.384765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19118925426155328,
+      "step": 6530
+    },
+    {
+      "epoch": 0.1635,
+      "grad_norm": 29.375,
+      "grad_norm_var": 8.820833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3298,
+      "loss/crossentropy": 2.119840921461582,
+      "loss/hidden": 3.39296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20046985391527414,
+      "step": 6540
+    },
+    {
+      "epoch": 0.16375,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.1910807291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4507,
+      "loss/crossentropy": 2.109931045770645,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18326662238687277,
+      "step": 6550
+    },
+    {
+      "epoch": 0.164,
+      "grad_norm": 35.25,
+      "grad_norm_var": 8.6625,
+      "learning_rate": 0.0001,
+      "loss": 7.43,
+      "loss/crossentropy": 2.090756069123745,
+      "loss/hidden": 3.348046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1922204466536641,
+      "step": 6560
+    },
+    {
+      "epoch": 0.16425,
+      "grad_norm": 30.625,
+      "grad_norm_var": 14.567643229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.1796,
+      "loss/crossentropy": 1.9266313910484314,
+      "loss/hidden": 3.465234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20681370329111814,
+      "step": 6570
+    },
+    {
+      "epoch": 0.1645,
+      "grad_norm": 37.5,
+      "grad_norm_var": 12.9447265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5097,
+      "loss/crossentropy": 1.875116103887558,
+      "loss/hidden": 3.49921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2045787101611495,
+      "step": 6580
+    },
+    {
+      "epoch": 0.16475,
+      "grad_norm": 29.875,
+      "grad_norm_var": 5.706184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.329,
+      "loss/crossentropy": 2.116366655379534,
+      "loss/hidden": 3.4546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18577212654054165,
+      "step": 6590
+    },
+    {
+      "epoch": 0.165,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.3650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3765,
+      "loss/crossentropy": 2.0037689693272114,
+      "loss/hidden": 3.38125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1971780034713447,
+      "step": 6600
+    },
+    {
+      "epoch": 0.16525,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.601497395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4001,
+      "loss/crossentropy": 2.1523181863129137,
+      "loss/hidden": 3.405859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.192273567058146,
+      "step": 6610
+    },
+    {
+      "epoch": 0.1655,
+      "grad_norm": 29.875,
+      "grad_norm_var": 7.6525390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3804,
+      "loss/crossentropy": 2.0919234342873096,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20757663380354643,
+      "step": 6620
+    },
+    {
+      "epoch": 0.16575,
+      "grad_norm": 31.875,
+      "grad_norm_var": 6.917708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3889,
+      "loss/crossentropy": 2.035097151994705,
+      "loss/hidden": 3.31484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19357213731855155,
+      "step": 6630
+    },
+    {
+      "epoch": 0.166,
+      "grad_norm": 28.125,
+      "grad_norm_var": 2.4607245906905574e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5149,
+      "loss/crossentropy": 2.114539227634668,
+      "loss/hidden": 3.358984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19631449952721597,
+      "step": 6640
+    },
+    {
+      "epoch": 0.16625,
+      "grad_norm": 28.375,
+      "grad_norm_var": 2.4607245908931773e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3157,
+      "loss/crossentropy": 2.0170676171779633,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1893145913258195,
+      "step": 6650
+    },
+    {
+      "epoch": 0.1665,
+      "grad_norm": 28.0,
+      "grad_norm_var": 32.25807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2947,
+      "loss/crossentropy": 1.9412188947200775,
+      "loss/hidden": 3.41953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18349691890180111,
+      "step": 6660
+    },
+    {
+      "epoch": 0.16675,
+      "grad_norm": 30.25,
+      "grad_norm_var": 48.2375,
+      "learning_rate": 0.0001,
+      "loss": 7.4086,
+      "loss/crossentropy": 2.1157006829977036,
+      "loss/hidden": 3.410546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1936411712318659,
+      "step": 6670
+    },
+    {
+      "epoch": 0.167,
+      "grad_norm": 29.75,
+      "grad_norm_var": 36.18333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4096,
+      "loss/crossentropy": 1.9384170174598694,
+      "loss/hidden": 3.46015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1831628430634737,
+      "step": 6680
+    },
+    {
+      "epoch": 0.16725,
+      "grad_norm": 30.0,
+      "grad_norm_var": 62.6150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4604,
+      "loss/crossentropy": 2.152873657643795,
+      "loss/hidden": 3.421484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20049556214362382,
+      "step": 6690
+    },
+    {
+      "epoch": 0.1675,
+      "grad_norm": 29.75,
+      "grad_norm_var": 28.671875,
+      "learning_rate": 0.0001,
+      "loss": 7.5739,
+      "loss/crossentropy": 2.1935679107904433,
+      "loss/hidden": 3.359765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20188184324651956,
+      "step": 6700
+    },
+    {
+      "epoch": 0.16775,
+      "grad_norm": 28.25,
+      "grad_norm_var": 2.278580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4709,
+      "loss/crossentropy": 2.062030902504921,
+      "loss/hidden": 3.4328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18874304387718438,
+      "step": 6710
+    },
+    {
+      "epoch": 0.168,
+      "grad_norm": 29.125,
+      "grad_norm_var": 3.3150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3655,
+      "loss/crossentropy": 1.999978879839182,
+      "loss/hidden": 3.3875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18526637642644345,
+      "step": 6720
+    },
+    {
+      "epoch": 0.16825,
+      "grad_norm": 35.25,
+      "grad_norm_var": 6.237239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4037,
+      "loss/crossentropy": 2.0561595499515533,
+      "loss/hidden": 3.41015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20046296287328005,
+      "step": 6730
+    },
+    {
+      "epoch": 0.1685,
+      "grad_norm": 28.75,
+      "grad_norm_var": 5.5619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.368,
+      "loss/crossentropy": 2.0664093092083933,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1885912848636508,
+      "step": 6740
+    },
+    {
+      "epoch": 0.16875,
+      "grad_norm": 30.875,
+      "grad_norm_var": 5.843489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4899,
+      "loss/crossentropy": 2.1205774366855623,
+      "loss/hidden": 3.491796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21412673257291318,
+      "step": 6750
+    },
+    {
+      "epoch": 0.169,
+      "grad_norm": 32.0,
+      "grad_norm_var": 6.3775390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3708,
+      "loss/crossentropy": 2.0314668610692026,
+      "loss/hidden": 3.33203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18550706487149,
+      "step": 6760
+    },
+    {
+      "epoch": 0.16925,
+      "grad_norm": 32.5,
+      "grad_norm_var": 7.378059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4055,
+      "loss/crossentropy": 2.1428691864013674,
+      "loss/hidden": 3.506640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21024896781891583,
+      "step": 6770
+    },
+    {
+      "epoch": 0.1695,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.9791666666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4444,
+      "loss/crossentropy": 1.9500044576823712,
+      "loss/hidden": 3.434375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19028451843187213,
+      "step": 6780
+    },
+    {
+      "epoch": 0.16975,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.2853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4369,
+      "loss/crossentropy": 2.1563921123743057,
+      "loss/hidden": 3.3328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18290520180016756,
+      "step": 6790
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 38.0,
+      "grad_norm_var": 8.0947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3756,
+      "loss/crossentropy": 2.133736363053322,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1851572971791029,
+      "step": 6800
+    },
+    {
+      "epoch": 0.17025,
+      "grad_norm": 34.75,
+      "grad_norm_var": 14.757747395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2181,
+      "loss/crossentropy": 2.054655596613884,
+      "loss/hidden": 3.349609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18349520340561867,
+      "step": 6810
+    },
+    {
+      "epoch": 0.1705,
+      "grad_norm": 30.375,
+      "grad_norm_var": 5.17890625,
+      "learning_rate": 0.0001,
+      "loss": 7.303,
+      "loss/crossentropy": 2.024763736128807,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19044207576662303,
+      "step": 6820
+    },
+    {
+      "epoch": 0.17075,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.7955729166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4884,
+      "loss/crossentropy": 1.9924081854522229,
+      "loss/hidden": 3.49609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20656490996479987,
+      "step": 6830
+    },
+    {
+      "epoch": 0.171,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.215559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4845,
+      "loss/crossentropy": 2.1104799427092074,
+      "loss/hidden": 3.334375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18316805781796575,
+      "step": 6840
+    },
+    {
+      "epoch": 0.17125,
+      "grad_norm": 31.625,
+      "grad_norm_var": 25.610872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3771,
+      "loss/crossentropy": 2.02793128117919,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17605492258444427,
+      "step": 6850
+    },
+    {
+      "epoch": 0.1715,
+      "grad_norm": 29.375,
+      "grad_norm_var": 34.992122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5157,
+      "loss/crossentropy": 2.1305345237255096,
+      "loss/hidden": 3.436328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19113806802779437,
+      "step": 6860
+    },
+    {
+      "epoch": 0.17175,
+      "grad_norm": 28.375,
+      "grad_norm_var": 12.389322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2752,
+      "loss/crossentropy": 2.0788576349616052,
+      "loss/hidden": 3.388671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19617441901937127,
+      "step": 6870
+    },
+    {
+      "epoch": 0.172,
+      "grad_norm": 27.625,
+      "grad_norm_var": 17.758072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4366,
+      "loss/crossentropy": 2.068412736058235,
+      "loss/hidden": 3.355078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18466003462672234,
+      "step": 6880
+    },
+    {
+      "epoch": 0.17225,
+      "grad_norm": 30.375,
+      "grad_norm_var": 19.836393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5047,
+      "loss/crossentropy": 2.07881121635437,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19641269743442535,
+      "step": 6890
+    },
+    {
+      "epoch": 0.1725,
+      "grad_norm": 29.625,
+      "grad_norm_var": 14.70625,
+      "learning_rate": 0.0001,
+      "loss": 7.4086,
+      "loss/crossentropy": 1.84702168405056,
+      "loss/hidden": 3.469140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1932983512058854,
+      "step": 6900
+    },
+    {
+      "epoch": 0.17275,
+      "grad_norm": 31.875,
+      "grad_norm_var": 31.006705729166665,
+      "learning_rate": 0.0001,
+      "loss": 7.3715,
+      "loss/crossentropy": 2.099086304008961,
+      "loss/hidden": 3.406640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20250021573156118,
+      "step": 6910
+    },
+    {
+      "epoch": 0.173,
+      "grad_norm": 29.0,
+      "grad_norm_var": 26.07265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3343,
+      "loss/crossentropy": 2.0957569405436516,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2010388659313321,
+      "step": 6920
+    },
+    {
+      "epoch": 0.17325,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.753059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4937,
+      "loss/crossentropy": 2.071177572757006,
+      "loss/hidden": 3.445703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19133195597678423,
+      "step": 6930
+    },
+    {
+      "epoch": 0.1735,
+      "grad_norm": 31.125,
+      "grad_norm_var": 5.88515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4757,
+      "loss/crossentropy": 2.0803056344389916,
+      "loss/hidden": 3.51875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20995833892375232,
+      "step": 6940
+    },
+    {
+      "epoch": 0.17375,
+      "grad_norm": 28.5,
+      "grad_norm_var": 7.328059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2293,
+      "loss/crossentropy": 1.9285863403230905,
+      "loss/hidden": 3.283203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1665677004493773,
+      "step": 6950
+    },
+    {
+      "epoch": 0.174,
+      "grad_norm": 27.375,
+      "grad_norm_var": 11.585872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.313,
+      "loss/crossentropy": 2.0258478805422784,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1821833540685475,
+      "step": 6960
+    },
+    {
+      "epoch": 0.17425,
+      "grad_norm": 30.5,
+      "grad_norm_var": 10.760416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5413,
+      "loss/crossentropy": 2.1308654129505156,
+      "loss/hidden": 3.43203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21524183861911297,
+      "step": 6970
+    },
+    {
+      "epoch": 0.1745,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.138541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4692,
+      "loss/crossentropy": 2.1182237058877944,
+      "loss/hidden": 3.362890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18946228343993426,
+      "step": 6980
+    },
+    {
+      "epoch": 0.17475,
+      "grad_norm": 31.0,
+      "grad_norm_var": 5.499934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4658,
+      "loss/crossentropy": 2.0863646306097507,
+      "loss/hidden": 3.38984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2054815970361233,
+      "step": 6990
+    },
+    {
+      "epoch": 0.175,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.7067057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3196,
+      "loss/crossentropy": 2.1002516582608224,
+      "loss/hidden": 3.40546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19273097421973945,
+      "step": 7000
+    },
+    {
+      "epoch": 0.17525,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.7999348958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3643,
+      "loss/crossentropy": 2.015849883854389,
+      "loss/hidden": 3.383984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18935495987534523,
+      "step": 7010
+    },
+    {
+      "epoch": 0.1755,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.3645833333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3973,
+      "loss/crossentropy": 2.072261115908623,
+      "loss/hidden": 3.482421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19815812185406684,
+      "step": 7020
+    },
+    {
+      "epoch": 0.17575,
+      "grad_norm": 29.375,
+      "grad_norm_var": 8.442708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4193,
+      "loss/crossentropy": 2.1367180705070496,
+      "loss/hidden": 3.31328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1970324844121933,
+      "step": 7030
+    },
+    {
+      "epoch": 0.176,
+      "grad_norm": 30.375,
+      "grad_norm_var": 5.426822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.518,
+      "loss/crossentropy": 2.210773140192032,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1955376474186778,
+      "step": 7040
+    },
+    {
+      "epoch": 0.17625,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.1791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3883,
+      "loss/crossentropy": 2.1343745410442354,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19158909022808074,
+      "step": 7050
+    },
+    {
+      "epoch": 0.1765,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.112434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4451,
+      "loss/crossentropy": 1.9646480686962604,
+      "loss/hidden": 3.465234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19925388041883707,
+      "step": 7060
+    },
+    {
+      "epoch": 0.17675,
+      "grad_norm": 31.5,
+      "grad_norm_var": 3.3275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4874,
+      "loss/crossentropy": 2.1882633604109287,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19836742132902146,
+      "step": 7070
+    },
+    {
+      "epoch": 0.177,
+      "grad_norm": 34.75,
+      "grad_norm_var": 5.7291015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4907,
+      "loss/crossentropy": 2.2362487465143204,
+      "loss/hidden": 3.33203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19153916742652655,
+      "step": 7080
+    },
+    {
+      "epoch": 0.17725,
+      "grad_norm": 29.5,
+      "grad_norm_var": 5.002083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4459,
+      "loss/crossentropy": 2.164398466050625,
+      "loss/hidden": 3.43828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20325577780604362,
+      "step": 7090
+    },
+    {
+      "epoch": 0.1775,
+      "grad_norm": 32.25,
+      "grad_norm_var": 21.282747395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4933,
+      "loss/crossentropy": 2.1975876331329345,
+      "loss/hidden": 3.455078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19638306740671396,
+      "step": 7100
+    },
+    {
+      "epoch": 0.17775,
+      "grad_norm": 28.5,
+      "grad_norm_var": 32.18020833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.358,
+      "loss/crossentropy": 1.9753169894218445,
+      "loss/hidden": 3.416015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17774544414132834,
+      "step": 7110
+    },
+    {
+      "epoch": 0.178,
+      "grad_norm": 30.0,
+      "grad_norm_var": 11.808333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4718,
+      "loss/crossentropy": 2.0591190218925477,
+      "loss/hidden": 3.4578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19963474106043577,
+      "step": 7120
+    },
+    {
+      "epoch": 0.17825,
+      "grad_norm": 29.25,
+      "grad_norm_var": 7.753580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3979,
+      "loss/crossentropy": 2.062809920310974,
+      "loss/hidden": 3.437109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19096513669937848,
+      "step": 7130
+    },
+    {
+      "epoch": 0.1785,
+      "grad_norm": 30.125,
+      "grad_norm_var": 6.6025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3036,
+      "loss/crossentropy": 2.0240534149110316,
+      "loss/hidden": 3.4453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18496394343674183,
+      "step": 7140
+    },
+    {
+      "epoch": 0.17875,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.4942057291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4778,
+      "loss/crossentropy": 2.124583348631859,
+      "loss/hidden": 3.44921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2003519142046571,
+      "step": 7150
+    },
+    {
+      "epoch": 0.179,
+      "grad_norm": 29.0,
+      "grad_norm_var": 13.8978515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3866,
+      "loss/crossentropy": 2.035899819433689,
+      "loss/hidden": 3.41484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18300745636224747,
+      "step": 7160
+    },
+    {
+      "epoch": 0.17925,
+      "grad_norm": 28.125,
+      "grad_norm_var": 18.375455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3499,
+      "loss/crossentropy": 2.086082286387682,
+      "loss/hidden": 3.26875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17294995756819845,
+      "step": 7170
+    },
+    {
+      "epoch": 0.1795,
+      "grad_norm": 28.5,
+      "grad_norm_var": 22.834830729166665,
+      "learning_rate": 0.0001,
+      "loss": 7.4265,
+      "loss/crossentropy": 2.0105697728693483,
+      "loss/hidden": 3.393359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19119318593293427,
+      "step": 7180
+    },
+    {
+      "epoch": 0.17975,
+      "grad_norm": 33.0,
+      "grad_norm_var": 19.51640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4722,
+      "loss/crossentropy": 2.1506593719124796,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19607089888304471,
+      "step": 7190
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 31.5,
+      "grad_norm_var": 9.854622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3205,
+      "loss/crossentropy": 2.0767677523195744,
+      "loss/hidden": 3.436328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19086614530533552,
+      "step": 7200
+    },
+    {
+      "epoch": 0.18025,
+      "grad_norm": 27.625,
+      "grad_norm_var": 8.885416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2908,
+      "loss/crossentropy": 2.230179136991501,
+      "loss/hidden": 3.26953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18500201255083085,
+      "step": 7210
+    },
+    {
+      "epoch": 0.1805,
+      "grad_norm": 38.25,
+      "grad_norm_var": 8.60625,
+      "learning_rate": 0.0001,
+      "loss": 7.4664,
+      "loss/crossentropy": 2.192136238515377,
+      "loss/hidden": 3.380859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1965922711417079,
+      "step": 7220
+    },
+    {
+      "epoch": 0.18075,
+      "grad_norm": 33.0,
+      "grad_norm_var": 7.258333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3796,
+      "loss/crossentropy": 2.010923378914595,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19027266185730696,
+      "step": 7230
+    },
+    {
+      "epoch": 0.181,
+      "grad_norm": 39.75,
+      "grad_norm_var": 11.3353515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5537,
+      "loss/crossentropy": 2.055556283891201,
+      "loss/hidden": 3.47421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19257053220644593,
+      "step": 7240
+    },
+    {
+      "epoch": 0.18125,
+      "grad_norm": 27.5,
+      "grad_norm_var": 14.424739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4216,
+      "loss/crossentropy": 2.136477355659008,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20195687096565962,
+      "step": 7250
+    },
+    {
+      "epoch": 0.1815,
+      "grad_norm": 30.125,
+      "grad_norm_var": 9.70390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4656,
+      "loss/crossentropy": 2.0134367659687995,
+      "loss/hidden": 3.523046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22354123163968326,
+      "step": 7260
+    },
+    {
+      "epoch": 0.18175,
+      "grad_norm": 30.0,
+      "grad_norm_var": 6.324739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3959,
+      "loss/crossentropy": 2.1245115220546724,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19102834183722733,
+      "step": 7270
+    },
+    {
+      "epoch": 0.182,
+      "grad_norm": 29.125,
+      "grad_norm_var": 18.903580729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3761,
+      "loss/crossentropy": 2.035867254436016,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19426564145833253,
+      "step": 7280
+    },
+    {
+      "epoch": 0.18225,
+      "grad_norm": 30.5,
+      "grad_norm_var": 20.406705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3884,
+      "loss/crossentropy": 1.9805133253335954,
+      "loss/hidden": 3.370703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19129701480269432,
+      "step": 7290
+    },
+    {
+      "epoch": 0.1825,
+      "grad_norm": 31.375,
+      "grad_norm_var": 5.842708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3833,
+      "loss/crossentropy": 2.0621849209070207,
+      "loss/hidden": 3.417578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19919742476195096,
+      "step": 7300
+    },
+    {
+      "epoch": 0.18275,
+      "grad_norm": 47.0,
+      "grad_norm_var": 2.675771470406222e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2357,
+      "loss/crossentropy": 2.1282688602805138,
+      "loss/hidden": 3.298046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18128441767767073,
+      "step": 7310
+    },
+    {
+      "epoch": 0.183,
+      "grad_norm": 29.25,
+      "grad_norm_var": 28.158333333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4531,
+      "loss/crossentropy": 2.1078659296035767,
+      "loss/hidden": 3.580078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2171280149370432,
+      "step": 7320
+    },
+    {
+      "epoch": 0.18325,
+      "grad_norm": 35.25,
+      "grad_norm_var": 25.365625,
+      "learning_rate": 0.0001,
+      "loss": 7.3538,
+      "loss/crossentropy": 2.200744313001633,
+      "loss/hidden": 3.37421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19498275145888327,
+      "step": 7330
+    },
+    {
+      "epoch": 0.1835,
+      "grad_norm": 28.375,
+      "grad_norm_var": 23.859830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3437,
+      "loss/crossentropy": 1.9968993581831456,
+      "loss/hidden": 3.3859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1875888810493052,
+      "step": 7340
+    },
+    {
+      "epoch": 0.18375,
+      "grad_norm": 28.0,
+      "grad_norm_var": 8.989518229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3984,
+      "loss/crossentropy": 2.0748091831803324,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1947355069220066,
+      "step": 7350
+    },
+    {
+      "epoch": 0.184,
+      "grad_norm": 29.875,
+      "grad_norm_var": 7.728125,
+      "learning_rate": 0.0001,
+      "loss": 7.3008,
+      "loss/crossentropy": 2.067877373099327,
+      "loss/hidden": 3.313671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18336500320583582,
+      "step": 7360
+    },
+    {
+      "epoch": 0.18425,
+      "grad_norm": 29.25,
+      "grad_norm_var": 13.189583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4547,
+      "loss/crossentropy": 2.152864509820938,
+      "loss/hidden": 3.451953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2102669222280383,
+      "step": 7370
+    },
+    {
+      "epoch": 0.1845,
+      "grad_norm": 28.5,
+      "grad_norm_var": 8.7556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.286,
+      "loss/crossentropy": 1.9991149730980395,
+      "loss/hidden": 3.327734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18364950213581324,
+      "step": 7380
+    },
+    {
+      "epoch": 0.18475,
+      "grad_norm": 31.75,
+      "grad_norm_var": 7.770572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3546,
+      "loss/crossentropy": 2.0513292245566843,
+      "loss/hidden": 3.355078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1970413200557232,
+      "step": 7390
+    },
+    {
+      "epoch": 0.185,
+      "grad_norm": 30.125,
+      "grad_norm_var": 7.620572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5116,
+      "loss/crossentropy": 2.102216296643019,
+      "loss/hidden": 3.51875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21091360161080958,
+      "step": 7400
+    },
+    {
+      "epoch": 0.18525,
+      "grad_norm": 32.75,
+      "grad_norm_var": 9.191666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3421,
+      "loss/crossentropy": 1.9926266744732857,
+      "loss/hidden": 3.455078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19832582902163268,
+      "step": 7410
+    },
+    {
+      "epoch": 0.1855,
+      "grad_norm": 29.5,
+      "grad_norm_var": 5.945833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4563,
+      "loss/crossentropy": 2.141331580281258,
+      "loss/hidden": 3.408203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20020943265408278,
+      "step": 7420
+    },
+    {
+      "epoch": 0.18575,
+      "grad_norm": 31.0,
+      "grad_norm_var": 7.198372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4227,
+      "loss/crossentropy": 1.9694693490862847,
+      "loss/hidden": 3.421484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18249646089971067,
+      "step": 7430
+    },
+    {
+      "epoch": 0.186,
+      "grad_norm": 29.375,
+      "grad_norm_var": 5.9697265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4119,
+      "loss/crossentropy": 2.1407265037298204,
+      "loss/hidden": 3.327734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17720893137156962,
+      "step": 7440
+    },
+    {
+      "epoch": 0.18625,
+      "grad_norm": 29.875,
+      "grad_norm_var": 0.8072916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3441,
+      "loss/crossentropy": 2.124198019504547,
+      "loss/hidden": 3.424609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1946978410705924,
+      "step": 7450
+    },
+    {
+      "epoch": 0.1865,
+      "grad_norm": 32.75,
+      "grad_norm_var": 1.9634765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4652,
+      "loss/crossentropy": 2.131994958221912,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19461573138833047,
+      "step": 7460
+    },
+    {
+      "epoch": 0.18675,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.7280598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3278,
+      "loss/crossentropy": 2.117748848348856,
+      "loss/hidden": 3.2578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18584235943853855,
+      "step": 7470
+    },
+    {
+      "epoch": 0.187,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.466080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5817,
+      "loss/crossentropy": 2.1364282086491584,
+      "loss/hidden": 3.460546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21069204956293106,
+      "step": 7480
+    },
+    {
+      "epoch": 0.18725,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.012434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5197,
+      "loss/crossentropy": 2.0523312032222747,
+      "loss/hidden": 3.50078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21221144162118435,
+      "step": 7490
+    },
+    {
+      "epoch": 0.1875,
+      "grad_norm": 33.75,
+      "grad_norm_var": 4.995768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4317,
+      "loss/crossentropy": 2.0852270901203154,
+      "loss/hidden": 3.359765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17908250950276852,
+      "step": 7500
+    },
+    {
+      "epoch": 0.18775,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.4518229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5444,
+      "loss/crossentropy": 2.059383874386549,
+      "loss/hidden": 3.436328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19440155941992998,
+      "step": 7510
+    },
+    {
+      "epoch": 0.188,
+      "grad_norm": 27.625,
+      "grad_norm_var": 2.851822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5199,
+      "loss/crossentropy": 2.1382876858115196,
+      "loss/hidden": 3.4078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19888029601424934,
+      "step": 7520
+    },
+    {
+      "epoch": 0.18825,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.073372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3012,
+      "loss/crossentropy": 2.0625696159899234,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18495072829537093,
+      "step": 7530
+    },
+    {
+      "epoch": 0.1885,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.2122395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3643,
+      "loss/crossentropy": 2.124967637658119,
+      "loss/hidden": 3.3328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18375679664313793,
+      "step": 7540
+    },
+    {
+      "epoch": 0.18875,
+      "grad_norm": 29.125,
+      "grad_norm_var": 3.2681640625,
+      "learning_rate": 0.0001,
+      "loss": 7.351,
+      "loss/crossentropy": 2.0680116668343542,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18827605471014977,
+      "step": 7550
+    },
+    {
+      "epoch": 0.189,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.5337890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4207,
+      "loss/crossentropy": 2.079096484184265,
+      "loss/hidden": 3.443359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20160295628011227,
+      "step": 7560
+    },
+    {
+      "epoch": 0.18925,
+      "grad_norm": 30.625,
+      "grad_norm_var": 18.2197265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4789,
+      "loss/crossentropy": 2.058067685365677,
+      "loss/hidden": 3.46640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21216327100992202,
+      "step": 7570
+    },
+    {
+      "epoch": 0.1895,
+      "grad_norm": 34.25,
+      "grad_norm_var": 14.415625,
+      "learning_rate": 0.0001,
+      "loss": 7.4987,
+      "loss/crossentropy": 2.0142914205789566,
+      "loss/hidden": 3.580078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20794902741909027,
+      "step": 7580
+    },
+    {
+      "epoch": 0.18975,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.9809895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4585,
+      "loss/crossentropy": 2.299562671780586,
+      "loss/hidden": 3.347265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19880922697484493,
+      "step": 7590
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 30.875,
+      "grad_norm_var": 15.101822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3204,
+      "loss/crossentropy": 2.1472302600741386,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18796155080199242,
+      "step": 7600
+    },
+    {
+      "epoch": 0.19025,
+      "grad_norm": 29.0,
+      "grad_norm_var": 2.5940733610451533e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5335,
+      "loss/crossentropy": 2.1664531916379928,
+      "loss/hidden": 3.3984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18874028734862805,
+      "step": 7610
+    },
+    {
+      "epoch": 0.1905,
+      "grad_norm": 29.125,
+      "grad_norm_var": 0.8843098958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4846,
+      "loss/crossentropy": 2.0765088513493537,
+      "loss/hidden": 3.34375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18674521408975125,
+      "step": 7620
+    },
+    {
+      "epoch": 0.19075,
+      "grad_norm": 27.25,
+      "grad_norm_var": 4.381184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3401,
+      "loss/crossentropy": 1.7713539503514766,
+      "loss/hidden": 3.515234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1749590938910842,
+      "step": 7630
+    },
+    {
+      "epoch": 0.191,
+      "grad_norm": 33.5,
+      "grad_norm_var": 5.26640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4838,
+      "loss/crossentropy": 1.922049730271101,
+      "loss/hidden": 3.388671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18183694053441285,
+      "step": 7640
+    },
+    {
+      "epoch": 0.19125,
+      "grad_norm": 30.5,
+      "grad_norm_var": 4.030989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3887,
+      "loss/crossentropy": 2.094516658782959,
+      "loss/hidden": 3.452734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18757299687713386,
+      "step": 7650
+    },
+    {
+      "epoch": 0.1915,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.5035807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4445,
+      "loss/crossentropy": 1.999229770898819,
+      "loss/hidden": 3.516015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19523975029587745,
+      "step": 7660
+    },
+    {
+      "epoch": 0.19175,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.0275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4626,
+      "loss/crossentropy": 2.0366951674222946,
+      "loss/hidden": 3.42265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18243371956050397,
+      "step": 7670
+    },
+    {
+      "epoch": 0.192,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.5931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5274,
+      "loss/crossentropy": 2.0489632681012155,
+      "loss/hidden": 3.564453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21230401135981083,
+      "step": 7680
+    },
+    {
+      "epoch": 0.19225,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.738997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4128,
+      "loss/crossentropy": 2.0826203912496566,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18644160348922015,
+      "step": 7690
+    },
+    {
+      "epoch": 0.1925,
+      "grad_norm": 29.875,
+      "grad_norm_var": 4.351041666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5195,
+      "loss/crossentropy": 2.1674430795013904,
+      "loss/hidden": 3.3640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19499621093273162,
+      "step": 7700
+    },
+    {
+      "epoch": 0.19275,
+      "grad_norm": 28.625,
+      "grad_norm_var": 4.939322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4236,
+      "loss/crossentropy": 2.092283549904823,
+      "loss/hidden": 3.39453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19506504610180855,
+      "step": 7710
+    },
+    {
+      "epoch": 0.193,
+      "grad_norm": 33.0,
+      "grad_norm_var": 9.080208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3501,
+      "loss/crossentropy": 2.0526101261377336,
+      "loss/hidden": 3.51328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19770189765840768,
+      "step": 7720
+    },
+    {
+      "epoch": 0.19325,
+      "grad_norm": 29.0,
+      "grad_norm_var": 9.731184895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4607,
+      "loss/crossentropy": 2.05912861302495,
+      "loss/hidden": 3.41328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19287334326654673,
+      "step": 7730
+    },
+    {
+      "epoch": 0.1935,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.8052083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4243,
+      "loss/crossentropy": 1.8983285859227181,
+      "loss/hidden": 3.4203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18626301139593124,
+      "step": 7740
+    },
+    {
+      "epoch": 0.19375,
+      "grad_norm": 40.75,
+      "grad_norm_var": 9.1541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4332,
+      "loss/crossentropy": 2.099681233614683,
+      "loss/hidden": 3.40546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20556394904851913,
+      "step": 7750
+    },
+    {
+      "epoch": 0.194,
+      "grad_norm": 37.0,
+      "grad_norm_var": 12.584375,
+      "learning_rate": 0.0001,
+      "loss": 7.516,
+      "loss/crossentropy": 2.0436215907335282,
+      "loss/hidden": 3.317578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17505232142284513,
+      "step": 7760
+    },
+    {
+      "epoch": 0.19425,
+      "grad_norm": 30.125,
+      "grad_norm_var": 5.2681640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5486,
+      "loss/crossentropy": 2.0449838273227217,
+      "loss/hidden": 3.46875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19324074545875192,
+      "step": 7770
+    },
+    {
+      "epoch": 0.1945,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.4518229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4052,
+      "loss/crossentropy": 2.1020638972520826,
+      "loss/hidden": 3.3234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1987349819391966,
+      "step": 7780
+    },
+    {
+      "epoch": 0.19475,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.525,
+      "learning_rate": 0.0001,
+      "loss": 7.3148,
+      "loss/crossentropy": 2.0913542471826077,
+      "loss/hidden": 3.38515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1900358498096466,
+      "step": 7790
+    },
+    {
+      "epoch": 0.195,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.2280598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.43,
+      "loss/crossentropy": 1.9448820307850838,
+      "loss/hidden": 3.423828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18228193083778024,
+      "step": 7800
+    },
+    {
+      "epoch": 0.19525,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.3791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4772,
+      "loss/crossentropy": 2.0547610491514208,
+      "loss/hidden": 3.459765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19370344914495946,
+      "step": 7810
+    },
+    {
+      "epoch": 0.1955,
+      "grad_norm": 31.375,
+      "grad_norm_var": 6.208072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4448,
+      "loss/crossentropy": 2.0798824220895766,
+      "loss/hidden": 3.431640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18972196318209172,
+      "step": 7820
+    },
+    {
+      "epoch": 0.19575,
+      "grad_norm": 29.5,
+      "grad_norm_var": 6.248893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5358,
+      "loss/crossentropy": 2.2324195951223373,
+      "loss/hidden": 3.4375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2047037549316883,
+      "step": 7830
+    },
+    {
+      "epoch": 0.196,
+      "grad_norm": 29.375,
+      "grad_norm_var": 4.453059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3114,
+      "loss/crossentropy": 2.1020479179918765,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19328910131007432,
+      "step": 7840
+    },
+    {
+      "epoch": 0.19625,
+      "grad_norm": 33.5,
+      "grad_norm_var": 590.0817057291666,
+      "learning_rate": 0.0001,
+      "loss": 7.4281,
+      "loss/crossentropy": 2.0953447744250298,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1854689259082079,
+      "step": 7850
+    },
+    {
+      "epoch": 0.1965,
+      "grad_norm": 33.0,
+      "grad_norm_var": 625.6192057291667,
+      "learning_rate": 0.0001,
+      "loss": 7.5283,
+      "loss/crossentropy": 2.061315707862377,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18506519980728625,
+      "step": 7860
+    },
+    {
+      "epoch": 0.19675,
+      "grad_norm": 32.0,
+      "grad_norm_var": 69.38430989583334,
+      "learning_rate": 0.0001,
+      "loss": 7.5314,
+      "loss/crossentropy": 2.1605025470256805,
+      "loss/hidden": 3.3890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20316088199615479,
+      "step": 7870
+    },
+    {
+      "epoch": 0.197,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.1067057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4244,
+      "loss/crossentropy": 2.1977868393063544,
+      "loss/hidden": 3.430078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1992826245725155,
+      "step": 7880
+    },
+    {
+      "epoch": 0.19725,
+      "grad_norm": 29.25,
+      "grad_norm_var": 9.578059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.502,
+      "loss/crossentropy": 2.0480964958667753,
+      "loss/hidden": 3.636328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21483200527727603,
+      "step": 7890
+    },
+    {
+      "epoch": 0.1975,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.6372395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.548,
+      "loss/crossentropy": 2.157261362671852,
+      "loss/hidden": 3.484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2084518164396286,
+      "step": 7900
+    },
+    {
+      "epoch": 0.19775,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.095572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3864,
+      "loss/crossentropy": 2.1441849052906035,
+      "loss/hidden": 3.414453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20142039898782968,
+      "step": 7910
+    },
+    {
+      "epoch": 0.198,
+      "grad_norm": 37.5,
+      "grad_norm_var": 16.170247395833332,
+      "learning_rate": 0.0001,
+      "loss": 7.429,
+      "loss/crossentropy": 2.001983726769686,
+      "loss/hidden": 3.448828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19066998092457652,
+      "step": 7920
+    },
+    {
+      "epoch": 0.19825,
+      "grad_norm": 29.375,
+      "grad_norm_var": 15.561393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4093,
+      "loss/crossentropy": 2.225774070620537,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1993710033595562,
+      "step": 7930
+    },
+    {
+      "epoch": 0.1985,
+      "grad_norm": 34.0,
+      "grad_norm_var": 25.343489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4932,
+      "loss/crossentropy": 2.0083594918251038,
+      "loss/hidden": 3.3453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19412665143609048,
+      "step": 7940
+    },
+    {
+      "epoch": 0.19875,
+      "grad_norm": 31.625,
+      "grad_norm_var": 23.877083333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.441,
+      "loss/crossentropy": 2.08128562271595,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17889103144407273,
+      "step": 7950
+    },
+    {
+      "epoch": 0.199,
+      "grad_norm": 27.875,
+      "grad_norm_var": 1.5921223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4251,
+      "loss/crossentropy": 2.1019147261977196,
+      "loss/hidden": 3.571484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2211546439677477,
+      "step": 7960
+    },
+    {
+      "epoch": 0.19925,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.3059895833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4213,
+      "loss/crossentropy": 2.14247687458992,
+      "loss/hidden": 3.420703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18870262056589127,
+      "step": 7970
+    },
+    {
+      "epoch": 0.1995,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.0947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5788,
+      "loss/crossentropy": 2.1896591186523438,
+      "loss/hidden": 3.479296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20883973222225904,
+      "step": 7980
+    },
+    {
+      "epoch": 0.19975,
+      "grad_norm": 28.25,
+      "grad_norm_var": 2.982291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5217,
+      "loss/crossentropy": 2.1884095311164855,
+      "loss/hidden": 3.44296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2022842913866043,
+      "step": 7990
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.379622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4975,
+      "loss/crossentropy": 2.225523295998573,
+      "loss/hidden": 3.4203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2032675376161933,
+      "step": 8000
+    },
+    {
+      "epoch": 0.20025,
+      "grad_norm": 30.625,
+      "grad_norm_var": 4.181705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3874,
+      "loss/crossentropy": 1.9566738605499268,
+      "loss/hidden": 3.576953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19702311754226684,
+      "step": 8010
+    },
+    {
+      "epoch": 0.2005,
+      "grad_norm": 32.75,
+      "grad_norm_var": 6.3509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4528,
+      "loss/crossentropy": 2.1517204724252226,
+      "loss/hidden": 3.466015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2005010774359107,
+      "step": 8020
+    },
+    {
+      "epoch": 0.20075,
+      "grad_norm": 30.875,
+      "grad_norm_var": 5.44140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5306,
+      "loss/crossentropy": 2.0300184957683087,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18887464031577111,
+      "step": 8030
+    },
+    {
+      "epoch": 0.201,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.1302083333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4025,
+      "loss/crossentropy": 2.0889772072434427,
+      "loss/hidden": 3.27109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17016669576987625,
+      "step": 8040
+    },
+    {
+      "epoch": 0.20125,
+      "grad_norm": 27.125,
+      "grad_norm_var": 94.78125,
+      "learning_rate": 0.0001,
+      "loss": 7.4379,
+      "loss/crossentropy": 2.158484524488449,
+      "loss/hidden": 3.410546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19672231934964657,
+      "step": 8050
+    },
+    {
+      "epoch": 0.2015,
+      "grad_norm": 38.0,
+      "grad_norm_var": 13.220572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4314,
+      "loss/crossentropy": 1.9623262777924537,
+      "loss/hidden": 3.339453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1833876773715019,
+      "step": 8060
+    },
+    {
+      "epoch": 0.20175,
+      "grad_norm": 29.875,
+      "grad_norm_var": 7.858268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4662,
+      "loss/crossentropy": 2.2177498638629913,
+      "loss/hidden": 3.383203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1954023003578186,
+      "step": 8070
+    },
+    {
+      "epoch": 0.202,
+      "grad_norm": 28.875,
+      "grad_norm_var": 7.627083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4792,
+      "loss/crossentropy": 2.1019868202507497,
+      "loss/hidden": 3.44375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1995641984976828,
+      "step": 8080
+    },
+    {
+      "epoch": 0.20225,
+      "grad_norm": 44.75,
+      "grad_norm_var": 20.4306640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5664,
+      "loss/crossentropy": 2.299755599349737,
+      "loss/hidden": 3.3015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19075682908296585,
+      "step": 8090
+    },
+    {
+      "epoch": 0.2025,
+      "grad_norm": 37.75,
+      "grad_norm_var": 2.5671221292944763e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4526,
+      "loss/crossentropy": 2.131952489167452,
+      "loss/hidden": 3.475390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1955398641526699,
+      "step": 8100
+    },
+    {
+      "epoch": 0.20275,
+      "grad_norm": 31.25,
+      "grad_norm_var": 20.342643229166665,
+      "learning_rate": 0.0001,
+      "loss": 7.4687,
+      "loss/crossentropy": 1.9825796701014042,
+      "loss/hidden": 3.475,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20226136669516565,
+      "step": 8110
+    },
+    {
+      "epoch": 0.203,
+      "grad_norm": 27.125,
+      "grad_norm_var": 15.192643229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.1604,
+      "loss/crossentropy": 2.0296560734510423,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1786259189248085,
+      "step": 8120
+    },
+    {
+      "epoch": 0.20325,
+      "grad_norm": 31.0,
+      "grad_norm_var": 7.1775390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2815,
+      "loss/crossentropy": 2.104996609687805,
+      "loss/hidden": 3.405859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19307580199092628,
+      "step": 8130
+    },
+    {
+      "epoch": 0.2035,
+      "grad_norm": 34.75,
+      "grad_norm_var": 6.820572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3774,
+      "loss/crossentropy": 2.1900447353720667,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20103423558175565,
+      "step": 8140
+    },
+    {
+      "epoch": 0.20375,
+      "grad_norm": 31.0,
+      "grad_norm_var": 5.48125,
+      "learning_rate": 0.0001,
+      "loss": 7.3628,
+      "loss/crossentropy": 2.0671978294849396,
+      "loss/hidden": 3.338671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18412660714238882,
+      "step": 8150
+    },
+    {
+      "epoch": 0.204,
+      "grad_norm": 41.25,
+      "grad_norm_var": 14.245572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5547,
+      "loss/crossentropy": 2.05537860840559,
+      "loss/hidden": 3.48515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1820721985772252,
+      "step": 8160
+    },
+    {
+      "epoch": 0.20425,
+      "grad_norm": 29.125,
+      "grad_norm_var": 13.8625,
+      "learning_rate": 0.0001,
+      "loss": 7.345,
+      "loss/crossentropy": 1.9336151838302613,
+      "loss/hidden": 3.564453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20152895338833332,
+      "step": 8170
+    },
+    {
+      "epoch": 0.2045,
+      "grad_norm": 28.125,
+      "grad_norm_var": 7.096875,
+      "learning_rate": 0.0001,
+      "loss": 7.5005,
+      "loss/crossentropy": 2.055295965075493,
+      "loss/hidden": 3.35234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1865659100934863,
+      "step": 8180
+    },
+    {
+      "epoch": 0.20475,
+      "grad_norm": 28.25,
+      "grad_norm_var": 8.805989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5324,
+      "loss/crossentropy": 1.9976115971803665,
+      "loss/hidden": 3.508203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2114253517240286,
+      "step": 8190
+    },
+    {
+      "epoch": 0.205,
+      "grad_norm": 29.875,
+      "grad_norm_var": 12.435416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4232,
+      "loss/crossentropy": 2.1122905567288397,
+      "loss/hidden": 3.400390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21649520397186278,
+      "step": 8200
+    },
+    {
+      "epoch": 0.20525,
+      "grad_norm": 28.125,
+      "grad_norm_var": 10.055208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4875,
+      "loss/crossentropy": 2.1132961876690386,
+      "loss/hidden": 3.47734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2026100393384695,
+      "step": 8210
+    },
+    {
+      "epoch": 0.2055,
+      "grad_norm": 29.125,
+      "grad_norm_var": 6.793489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3834,
+      "loss/crossentropy": 2.191230720281601,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19885572660714387,
+      "step": 8220
+    },
+    {
+      "epoch": 0.20575,
+      "grad_norm": 30.75,
+      "grad_norm_var": 4.630208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4544,
+      "loss/crossentropy": 2.080083931982517,
+      "loss/hidden": 3.474609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2282587742432952,
+      "step": 8230
+    },
+    {
+      "epoch": 0.206,
+      "grad_norm": 31.375,
+      "grad_norm_var": 11.458268229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.449,
+      "loss/crossentropy": 2.1384637162089346,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1842938730493188,
+      "step": 8240
+    },
+    {
+      "epoch": 0.20625,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.7087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3946,
+      "loss/crossentropy": 2.1355771869421005,
+      "loss/hidden": 3.264453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18052869867533444,
+      "step": 8250
+    },
+    {
+      "epoch": 0.2065,
+      "grad_norm": 31.75,
+      "grad_norm_var": 13.0447265625,
+      "learning_rate": 0.0001,
+      "loss": 7.6316,
+      "loss/crossentropy": 2.0203320410102608,
+      "loss/hidden": 3.445703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.192235934920609,
+      "step": 8260
+    },
+    {
+      "epoch": 0.20675,
+      "grad_norm": 29.5,
+      "grad_norm_var": 12.270247395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4572,
+      "loss/crossentropy": 1.9869592547416688,
+      "loss/hidden": 3.44140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.202651490829885,
+      "step": 8270
+    },
+    {
+      "epoch": 0.207,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.4822916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5227,
+      "loss/crossentropy": 2.051181730628014,
+      "loss/hidden": 3.43984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20994498692452906,
+      "step": 8280
+    },
+    {
+      "epoch": 0.20725,
+      "grad_norm": 43.75,
+      "grad_norm_var": 12.776822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3664,
+      "loss/crossentropy": 1.9618318520486355,
+      "loss/hidden": 3.558984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19053993374109268,
+      "step": 8290
+    },
+    {
+      "epoch": 0.2075,
+      "grad_norm": 30.5,
+      "grad_norm_var": 14.92890625,
+      "learning_rate": 0.0001,
+      "loss": 7.318,
+      "loss/crossentropy": 2.0114831268787383,
+      "loss/hidden": 3.459375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20329152811318635,
+      "step": 8300
+    },
+    {
+      "epoch": 0.20775,
+      "grad_norm": 34.0,
+      "grad_norm_var": 2.8869140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5489,
+      "loss/crossentropy": 2.1187786638736723,
+      "loss/hidden": 3.512890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20839224103838205,
+      "step": 8310
+    },
+    {
+      "epoch": 0.208,
+      "grad_norm": 27.875,
+      "grad_norm_var": 2.565559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3443,
+      "loss/crossentropy": 2.1263110756874086,
+      "loss/hidden": 3.32109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18627664018422366,
+      "step": 8320
+    },
+    {
+      "epoch": 0.20825,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.755989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3857,
+      "loss/crossentropy": 1.944861602783203,
+      "loss/hidden": 3.468359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18570939563214778,
+      "step": 8330
+    },
+    {
+      "epoch": 0.2085,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.340625,
+      "learning_rate": 0.0001,
+      "loss": 7.5344,
+      "loss/crossentropy": 2.1811724051833155,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19298948515206577,
+      "step": 8340
+    },
+    {
+      "epoch": 0.20875,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.56015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3826,
+      "loss/crossentropy": 2.152976579964161,
+      "loss/hidden": 3.305859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.190250195749104,
+      "step": 8350
+    },
+    {
+      "epoch": 0.209,
+      "grad_norm": 28.875,
+      "grad_norm_var": 1.3416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.323,
+      "loss/crossentropy": 2.2053099036216737,
+      "loss/hidden": 3.344140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1929330924525857,
+      "step": 8360
+    },
+    {
+      "epoch": 0.20925,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.65625,
+      "learning_rate": 0.0001,
+      "loss": 7.4376,
+      "loss/crossentropy": 2.011850906908512,
+      "loss/hidden": 3.442578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20794765576720237,
+      "step": 8370
+    },
+    {
+      "epoch": 0.2095,
+      "grad_norm": 29.25,
+      "grad_norm_var": 2.381705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4518,
+      "loss/crossentropy": 2.284806078672409,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2203810729086399,
+      "step": 8380
+    },
+    {
+      "epoch": 0.20975,
+      "grad_norm": 43.25,
+      "grad_norm_var": 17.695247395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.418,
+      "loss/crossentropy": 2.1161764934659004,
+      "loss/hidden": 3.475390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1863908626139164,
+      "step": 8390
+    },
+    {
+      "epoch": 0.21,
+      "grad_norm": 29.625,
+      "grad_norm_var": 19.8697265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4564,
+      "loss/crossentropy": 2.0293585821986198,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1884205201640725,
+      "step": 8400
+    },
+    {
+      "epoch": 0.21025,
+      "grad_norm": 27.25,
+      "grad_norm_var": 11.422330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4317,
+      "loss/crossentropy": 2.2351802065968513,
+      "loss/hidden": 3.376953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20424611177295446,
+      "step": 8410
+    },
+    {
+      "epoch": 0.2105,
+      "grad_norm": 30.125,
+      "grad_norm_var": 10.734830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3989,
+      "loss/crossentropy": 2.1349810734391212,
+      "loss/hidden": 3.3828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18979013338685036,
+      "step": 8420
+    },
+    {
+      "epoch": 0.21075,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.387955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3957,
+      "loss/crossentropy": 2.080636392533779,
+      "loss/hidden": 3.371875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18893024744465947,
+      "step": 8430
+    },
+    {
+      "epoch": 0.211,
+      "grad_norm": 28.75,
+      "grad_norm_var": 20.089518229166668,
+      "learning_rate": 0.0001,
+      "loss": 7.4847,
+      "loss/crossentropy": 1.9267802774906158,
+      "loss/hidden": 3.31484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18156335428357123,
+      "step": 8440
+    },
+    {
+      "epoch": 0.21125,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.2650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3839,
+      "loss/crossentropy": 2.14048397988081,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1889717074111104,
+      "step": 8450
+    },
+    {
+      "epoch": 0.2115,
+      "grad_norm": 28.375,
+      "grad_norm_var": 3.283124099188418e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4661,
+      "loss/crossentropy": 2.1406930878758432,
+      "loss/hidden": 3.699609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20904620084911585,
+      "step": 8460
+    },
+    {
+      "epoch": 0.21175,
+      "grad_norm": 30.75,
+      "grad_norm_var": 23.509830729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2867,
+      "loss/crossentropy": 2.1153231114149094,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18998019583523273,
+      "step": 8470
+    },
+    {
+      "epoch": 0.212,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.3139973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3997,
+      "loss/crossentropy": 2.169650764763355,
+      "loss/hidden": 3.276953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17556187361478806,
+      "step": 8480
+    },
+    {
+      "epoch": 0.21225,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.312239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4486,
+      "loss/crossentropy": 2.128412726521492,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1975807584822178,
+      "step": 8490
+    },
+    {
+      "epoch": 0.2125,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.903125,
+      "learning_rate": 0.0001,
+      "loss": 7.4526,
+      "loss/crossentropy": 2.029564914107323,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19610330546274782,
+      "step": 8500
+    },
+    {
+      "epoch": 0.21275,
+      "grad_norm": 31.375,
+      "grad_norm_var": 4.004622395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2845,
+      "loss/crossentropy": 1.9631854377686977,
+      "loss/hidden": 3.409765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1855375848710537,
+      "step": 8510
+    },
+    {
+      "epoch": 0.213,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.476822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4124,
+      "loss/crossentropy": 2.1494273841381073,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19540442936122418,
+      "step": 8520
+    },
+    {
+      "epoch": 0.21325,
+      "grad_norm": 31.625,
+      "grad_norm_var": 0.9452473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4659,
+      "loss/crossentropy": 2.0508621491491796,
+      "loss/hidden": 3.3078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18023168351501226,
+      "step": 8530
+    },
+    {
+      "epoch": 0.2135,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.5809895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4522,
+      "loss/crossentropy": 2.090715576708317,
+      "loss/hidden": 3.3875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18382459450513125,
+      "step": 8540
+    },
+    {
+      "epoch": 0.21375,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.986168000807314e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4917,
+      "loss/crossentropy": 2.177123633027077,
+      "loss/hidden": 3.359765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.197673611715436,
+      "step": 8550
+    },
+    {
+      "epoch": 0.214,
+      "grad_norm": 27.75,
+      "grad_norm_var": 4.035872395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3008,
+      "loss/crossentropy": 2.028589369356632,
+      "loss/hidden": 3.383984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.196718043461442,
+      "step": 8560
+    },
+    {
+      "epoch": 0.21425,
+      "grad_norm": 30.5,
+      "grad_norm_var": 4.758072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3829,
+      "loss/crossentropy": 2.02908306196332,
+      "loss/hidden": 3.3703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1757409404963255,
+      "step": 8570
+    },
+    {
+      "epoch": 0.2145,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.9754557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4107,
+      "loss/crossentropy": 2.0411842301487924,
+      "loss/hidden": 3.580078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2309743857011199,
+      "step": 8580
+    },
+    {
+      "epoch": 0.21475,
+      "grad_norm": 30.0,
+      "grad_norm_var": 25.7244140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4038,
+      "loss/crossentropy": 2.1026074662804604,
+      "loss/hidden": 3.490234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.202506691403687,
+      "step": 8590
+    },
+    {
+      "epoch": 0.215,
+      "grad_norm": 28.625,
+      "grad_norm_var": 3.386458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3559,
+      "loss/crossentropy": 2.1690615713596344,
+      "loss/hidden": 3.388671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19095612335950135,
+      "step": 8600
+    },
+    {
+      "epoch": 0.21525,
+      "grad_norm": 29.0,
+      "grad_norm_var": 23.880989583333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4164,
+      "loss/crossentropy": 2.099227898567915,
+      "loss/hidden": 3.376953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19162636022083462,
+      "step": 8610
+    },
+    {
+      "epoch": 0.2155,
+      "grad_norm": 34.25,
+      "grad_norm_var": 23.880208333333332,
+      "learning_rate": 0.0001,
+      "loss": 7.3725,
+      "loss/crossentropy": 1.9689884655177594,
+      "loss/hidden": 3.52265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1830376190133393,
+      "step": 8620
+    },
+    {
+      "epoch": 0.21575,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.77890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4511,
+      "loss/crossentropy": 2.0263702854514123,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18029460608959197,
+      "step": 8630
+    },
+    {
+      "epoch": 0.216,
+      "grad_norm": 32.0,
+      "grad_norm_var": 23.001822916666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3863,
+      "loss/crossentropy": 1.9046964697539805,
+      "loss/hidden": 3.44140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1935629203915596,
+      "step": 8640
+    },
+    {
+      "epoch": 0.21625,
+      "grad_norm": 33.5,
+      "grad_norm_var": 25.009375,
+      "learning_rate": 0.0001,
+      "loss": 7.3083,
+      "loss/crossentropy": 2.129426471889019,
+      "loss/hidden": 3.448046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20037918202579022,
+      "step": 8650
+    },
+    {
+      "epoch": 0.2165,
+      "grad_norm": 28.375,
+      "grad_norm_var": 12.242643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4603,
+      "loss/crossentropy": 2.2266604125499727,
+      "loss/hidden": 3.33203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18952292017638683,
+      "step": 8660
+    },
+    {
+      "epoch": 0.21675,
+      "grad_norm": 34.75,
+      "grad_norm_var": 19.517708333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4031,
+      "loss/crossentropy": 1.9756697475910188,
+      "loss/hidden": 3.529296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20053059812635182,
+      "step": 8670
+    },
+    {
+      "epoch": 0.217,
+      "grad_norm": 29.25,
+      "grad_norm_var": 8.315625,
+      "learning_rate": 0.0001,
+      "loss": 7.4013,
+      "loss/crossentropy": 2.13729098290205,
+      "loss/hidden": 3.3859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19440573658794164,
+      "step": 8680
+    },
+    {
+      "epoch": 0.21725,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.9936848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2982,
+      "loss/crossentropy": 2.0935462579131126,
+      "loss/hidden": 3.453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.200434254668653,
+      "step": 8690
+    },
+    {
+      "epoch": 0.2175,
+      "grad_norm": 31.625,
+      "grad_norm_var": 26.66640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4596,
+      "loss/crossentropy": 2.0081637501716614,
+      "loss/hidden": 3.42734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.196201959438622,
+      "step": 8700
+    },
+    {
+      "epoch": 0.21775,
+      "grad_norm": 28.125,
+      "grad_norm_var": 1.869627142435242e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4726,
+      "loss/crossentropy": 2.057238683104515,
+      "loss/hidden": 3.372265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18527965154498816,
+      "step": 8710
+    },
+    {
+      "epoch": 0.218,
+      "grad_norm": 30.25,
+      "grad_norm_var": 4.806705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.424,
+      "loss/crossentropy": 2.225848586857319,
+      "loss/hidden": 3.4,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20839223694056272,
+      "step": 8720
+    },
+    {
+      "epoch": 0.21825,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.2572265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3713,
+      "loss/crossentropy": 1.9737806752324105,
+      "loss/hidden": 3.413671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18166052605956792,
+      "step": 8730
+    },
+    {
+      "epoch": 0.2185,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.49765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3192,
+      "loss/crossentropy": 2.1625877559185027,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.196309875510633,
+      "step": 8740
+    },
+    {
+      "epoch": 0.21875,
+      "grad_norm": 31.625,
+      "grad_norm_var": 15.733333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3994,
+      "loss/crossentropy": 2.1306996777653695,
+      "loss/hidden": 3.3328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17798179090023042,
+      "step": 8750
+    },
+    {
+      "epoch": 0.219,
+      "grad_norm": 31.25,
+      "grad_norm_var": 5.465625,
+      "learning_rate": 0.0001,
+      "loss": 7.4225,
+      "loss/crossentropy": 2.168031161278486,
+      "loss/hidden": 3.423046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19842574130743743,
+      "step": 8760
+    },
+    {
+      "epoch": 0.21925,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.457747395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4515,
+      "loss/crossentropy": 2.081401216983795,
+      "loss/hidden": 3.3875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1922367751598358,
+      "step": 8770
+    },
+    {
+      "epoch": 0.2195,
+      "grad_norm": 28.625,
+      "grad_norm_var": 2.9389973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4641,
+      "loss/crossentropy": 1.9896476596593857,
+      "loss/hidden": 3.367578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18900877684354783,
+      "step": 8780
+    },
+    {
+      "epoch": 0.21975,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.5931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5034,
+      "loss/crossentropy": 2.131504286080599,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1981994620524347,
+      "step": 8790
+    },
+    {
+      "epoch": 0.22,
+      "grad_norm": 29.75,
+      "grad_norm_var": 17.612955729166668,
+      "learning_rate": 0.0001,
+      "loss": 7.3602,
+      "loss/crossentropy": 2.367118790745735,
+      "loss/hidden": 3.357421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20705808699131012,
+      "step": 8800
+    },
+    {
+      "epoch": 0.22025,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.588997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3512,
+      "loss/crossentropy": 1.915165586769581,
+      "loss/hidden": 3.41328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18030493911355733,
+      "step": 8810
+    },
+    {
+      "epoch": 0.2205,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.732291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3832,
+      "loss/crossentropy": 2.1191729307174683,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20246538575738668,
+      "step": 8820
+    },
+    {
+      "epoch": 0.22075,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.034309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4184,
+      "loss/crossentropy": 2.280582541972399,
+      "loss/hidden": 3.355078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1913912059739232,
+      "step": 8830
+    },
+    {
+      "epoch": 0.221,
+      "grad_norm": 28.75,
+      "grad_norm_var": 1.4629557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5177,
+      "loss/crossentropy": 2.07154730707407,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18417379464954137,
+      "step": 8840
+    },
+    {
+      "epoch": 0.22125,
+      "grad_norm": 29.25,
+      "grad_norm_var": 2.0306640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4098,
+      "loss/crossentropy": 2.0918928742408753,
+      "loss/hidden": 3.478515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21339783817529678,
+      "step": 8850
+    },
+    {
+      "epoch": 0.2215,
+      "grad_norm": 34.75,
+      "grad_norm_var": 2.4098307291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3907,
+      "loss/crossentropy": 2.0262165658175944,
+      "loss/hidden": 3.4796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1963033676147461,
+      "step": 8860
+    },
+    {
+      "epoch": 0.22175,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.1041015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3643,
+      "loss/crossentropy": 2.119324280321598,
+      "loss/hidden": 3.421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19243048634380103,
+      "step": 8870
+    },
+    {
+      "epoch": 0.222,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.3593098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.335,
+      "loss/crossentropy": 2.1042064100503923,
+      "loss/hidden": 3.43671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2055317424237728,
+      "step": 8880
+    },
+    {
+      "epoch": 0.22225,
+      "grad_norm": 29.0,
+      "grad_norm_var": 4.35390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3974,
+      "loss/crossentropy": 2.110988216102123,
+      "loss/hidden": 3.344140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18536690715700388,
+      "step": 8890
+    },
+    {
+      "epoch": 0.2225,
+      "grad_norm": 30.0,
+      "grad_norm_var": 6.014322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4155,
+      "loss/crossentropy": 2.033397987484932,
+      "loss/hidden": 3.367578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19111265633255242,
+      "step": 8900
+    },
+    {
+      "epoch": 0.22275,
+      "grad_norm": 32.25,
+      "grad_norm_var": 4.747916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4631,
+      "loss/crossentropy": 2.090746468305588,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20997797157615422,
+      "step": 8910
+    },
+    {
+      "epoch": 0.223,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.0247395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3647,
+      "loss/crossentropy": 2.107650229334831,
+      "loss/hidden": 3.41171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18095682561397552,
+      "step": 8920
+    },
+    {
+      "epoch": 0.22325,
+      "grad_norm": 33.0,
+      "grad_norm_var": 4.983072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4161,
+      "loss/crossentropy": 2.1033721581101417,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1837721960619092,
+      "step": 8930
+    },
+    {
+      "epoch": 0.2235,
+      "grad_norm": 30.0,
+      "grad_norm_var": 5.160872395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.381,
+      "loss/crossentropy": 2.2039036631584166,
+      "loss/hidden": 3.41953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20040026511996983,
+      "step": 8940
+    },
+    {
+      "epoch": 0.22375,
+      "grad_norm": 32.75,
+      "grad_norm_var": 5.625455729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4581,
+      "loss/crossentropy": 2.0015091970562935,
+      "loss/hidden": 3.4046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1968998895958066,
+      "step": 8950
+    },
+    {
+      "epoch": 0.224,
+      "grad_norm": 30.125,
+      "grad_norm_var": 7.976497395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.446,
+      "loss/crossentropy": 2.1770635031163694,
+      "loss/hidden": 3.290234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18164771795272827,
+      "step": 8960
+    },
+    {
+      "epoch": 0.22425,
+      "grad_norm": 26.875,
+      "grad_norm_var": 2.037239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2677,
+      "loss/crossentropy": 1.9846746385097505,
+      "loss/hidden": 3.45859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19175144601613284,
+      "step": 8970
+    },
+    {
+      "epoch": 0.2245,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.8353515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4645,
+      "loss/crossentropy": 2.169223573803902,
+      "loss/hidden": 3.33359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18991702441126107,
+      "step": 8980
+    },
+    {
+      "epoch": 0.22475,
+      "grad_norm": 29.875,
+      "grad_norm_var": 10.3228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4607,
+      "loss/crossentropy": 2.1323930069804193,
+      "loss/hidden": 3.348046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18631141390651465,
+      "step": 8990
+    },
+    {
+      "epoch": 0.225,
+      "grad_norm": 29.625,
+      "grad_norm_var": 11.431705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4203,
+      "loss/crossentropy": 2.1948125064373016,
+      "loss/hidden": 3.420703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21075339019298553,
+      "step": 9000
+    },
+    {
+      "epoch": 0.22525,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.21875,
+      "learning_rate": 0.0001,
+      "loss": 7.3937,
+      "loss/crossentropy": 2.061431697010994,
+      "loss/hidden": 3.357421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20084240343421697,
+      "step": 9010
+    },
+    {
+      "epoch": 0.2255,
+      "grad_norm": 30.0,
+      "grad_norm_var": 9.7125,
+      "learning_rate": 0.0001,
+      "loss": 7.5364,
+      "loss/crossentropy": 2.1422473564743996,
+      "loss/hidden": 3.38359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20135180205106734,
+      "step": 9020
+    },
+    {
+      "epoch": 0.22575,
+      "grad_norm": 29.0,
+      "grad_norm_var": 10.320247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3638,
+      "loss/crossentropy": 2.064805781841278,
+      "loss/hidden": 3.509375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19638751186430453,
+      "step": 9030
+    },
+    {
+      "epoch": 0.226,
+      "grad_norm": 29.0,
+      "grad_norm_var": 7.7306640625,
+      "learning_rate": 0.0001,
+      "loss": 7.367,
+      "loss/crossentropy": 2.0203323513269424,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1730576554313302,
+      "step": 9040
+    },
+    {
+      "epoch": 0.22625,
+      "grad_norm": 28.0,
+      "grad_norm_var": 7.5291015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4083,
+      "loss/crossentropy": 2.1449467122554777,
+      "loss/hidden": 3.428125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20745128113776445,
+      "step": 9050
+    },
+    {
+      "epoch": 0.2265,
+      "grad_norm": 29.625,
+      "grad_norm_var": 10.059830729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3739,
+      "loss/crossentropy": 2.277996188402176,
+      "loss/hidden": 3.359765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1961175424978137,
+      "step": 9060
+    },
+    {
+      "epoch": 0.22675,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.9497395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.425,
+      "loss/crossentropy": 2.128904873877764,
+      "loss/hidden": 3.307421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1912422338500619,
+      "step": 9070
+    },
+    {
+      "epoch": 0.227,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.4291666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3724,
+      "loss/crossentropy": 1.8966167330741883,
+      "loss/hidden": 3.315234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1772749178111553,
+      "step": 9080
+    },
+    {
+      "epoch": 0.22725,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.2955729166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4341,
+      "loss/crossentropy": 1.9993775576353072,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17548400331288577,
+      "step": 9090
+    },
+    {
+      "epoch": 0.2275,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.3212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4512,
+      "loss/crossentropy": 2.1553252935409546,
+      "loss/hidden": 3.41953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1904754728078842,
+      "step": 9100
+    },
+    {
+      "epoch": 0.22775,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.8264973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4299,
+      "loss/crossentropy": 2.0791067980229854,
+      "loss/hidden": 3.384765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18264174591749907,
+      "step": 9110
+    },
+    {
+      "epoch": 0.228,
+      "grad_norm": 30.625,
+      "grad_norm_var": 8.670572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3621,
+      "loss/crossentropy": 2.2495081633329392,
+      "loss/hidden": 3.3703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19295338317751884,
+      "step": 9120
+    },
+    {
+      "epoch": 0.22825,
+      "grad_norm": 34.25,
+      "grad_norm_var": 9.628580729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4069,
+      "loss/crossentropy": 2.0221784450113773,
+      "loss/hidden": 3.43046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18131311442703008,
+      "step": 9130
+    },
+    {
+      "epoch": 0.2285,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.8372395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3788,
+      "loss/crossentropy": 2.096472094208002,
+      "loss/hidden": 3.533984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19181215222924947,
+      "step": 9140
+    },
+    {
+      "epoch": 0.22875,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.9369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4625,
+      "loss/crossentropy": 2.161876367032528,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19645992666482925,
+      "step": 9150
+    },
+    {
+      "epoch": 0.229,
+      "grad_norm": 36.5,
+      "grad_norm_var": 13.144791666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3883,
+      "loss/crossentropy": 1.870260328054428,
+      "loss/hidden": 3.421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19089705124497414,
+      "step": 9160
+    },
+    {
+      "epoch": 0.22925,
+      "grad_norm": 30.125,
+      "grad_norm_var": 11.746809895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3778,
+      "loss/crossentropy": 2.304895442724228,
+      "loss/hidden": 3.30390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19963221047073604,
+      "step": 9170
+    },
+    {
+      "epoch": 0.2295,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.64839803006287e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4899,
+      "loss/crossentropy": 2.1439118653535845,
+      "loss/hidden": 3.46484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20680125001817942,
+      "step": 9180
+    },
+    {
+      "epoch": 0.22975,
+      "grad_norm": 29.625,
+      "grad_norm_var": 16.839518229166668,
+      "learning_rate": 0.0001,
+      "loss": 7.3153,
+      "loss/crossentropy": 1.9939407154917717,
+      "loss/hidden": 3.4125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19113040501251816,
+      "step": 9190
+    },
+    {
+      "epoch": 0.23,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.8309895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5357,
+      "loss/crossentropy": 2.088300883769989,
+      "loss/hidden": 3.476953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19538584928959607,
+      "step": 9200
+    },
+    {
+      "epoch": 0.23025,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.66640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3255,
+      "loss/crossentropy": 2.0176690459251403,
+      "loss/hidden": 3.361328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1823873495683074,
+      "step": 9210
+    },
+    {
+      "epoch": 0.2305,
+      "grad_norm": 32.75,
+      "grad_norm_var": 4.837955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3334,
+      "loss/crossentropy": 2.061754436790943,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18769590836018324,
+      "step": 9220
+    },
+    {
+      "epoch": 0.23075,
+      "grad_norm": 31.75,
+      "grad_norm_var": 4.683268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3932,
+      "loss/crossentropy": 2.0443666532635687,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19146692994982004,
+      "step": 9230
+    },
+    {
+      "epoch": 0.231,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.594073359575469e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5992,
+      "loss/crossentropy": 2.189889648556709,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19348510541021824,
+      "step": 9240
+    },
+    {
+      "epoch": 0.23125,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.594073359702976e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4217,
+      "loss/crossentropy": 2.0103170931339265,
+      "loss/hidden": 3.44140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19076041504740715,
+      "step": 9250
+    },
+    {
+      "epoch": 0.2315,
+      "grad_norm": 31.375,
+      "grad_norm_var": 13.5134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3004,
+      "loss/crossentropy": 2.0597189858555796,
+      "loss/hidden": 3.3421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18948603458702565,
+      "step": 9260
+    },
+    {
+      "epoch": 0.23175,
+      "grad_norm": 31.625,
+      "grad_norm_var": 13.402083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3946,
+      "loss/crossentropy": 2.0123729363083838,
+      "loss/hidden": 3.3828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1885578565299511,
+      "step": 9270
+    },
+    {
+      "epoch": 0.232,
+      "grad_norm": 29.875,
+      "grad_norm_var": 5.879622395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.1578,
+      "loss/crossentropy": 2.2501363843679427,
+      "loss/hidden": 3.286328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19268405642360448,
+      "step": 9280
+    },
+    {
+      "epoch": 0.23225,
+      "grad_norm": 30.25,
+      "grad_norm_var": 16.982291666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3481,
+      "loss/crossentropy": 2.076053886115551,
+      "loss/hidden": 3.571875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21077420487999915,
+      "step": 9290
+    },
+    {
+      "epoch": 0.2325,
+      "grad_norm": 29.5,
+      "grad_norm_var": 21.769791666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2711,
+      "loss/crossentropy": 2.193411388993263,
+      "loss/hidden": 3.33828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18807493168860673,
+      "step": 9300
+    },
+    {
+      "epoch": 0.23275,
+      "grad_norm": 34.25,
+      "grad_norm_var": 12.355143229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3819,
+      "loss/crossentropy": 2.166665832698345,
+      "loss/hidden": 3.3203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18620893750339745,
+      "step": 9310
+    },
+    {
+      "epoch": 0.233,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.002083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4076,
+      "loss/crossentropy": 2.022654353827238,
+      "loss/hidden": 3.2984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18289813362061977,
+      "step": 9320
+    },
+    {
+      "epoch": 0.23325,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.0509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3206,
+      "loss/crossentropy": 2.020027980953455,
+      "loss/hidden": 3.425390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18972874553874136,
+      "step": 9330
+    },
+    {
+      "epoch": 0.2335,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.098893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4091,
+      "loss/crossentropy": 2.029365235567093,
+      "loss/hidden": 3.43125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2005317559465766,
+      "step": 9340
+    },
+    {
+      "epoch": 0.23375,
+      "grad_norm": 29.5,
+      "grad_norm_var": 10.520768229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3704,
+      "loss/crossentropy": 1.95634398534894,
+      "loss/hidden": 3.433203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18962019477039577,
+      "step": 9350
+    },
+    {
+      "epoch": 0.234,
+      "grad_norm": 27.75,
+      "grad_norm_var": 13.475,
+      "learning_rate": 0.0001,
+      "loss": 7.3797,
+      "loss/crossentropy": 2.063827896118164,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19280518041923642,
+      "step": 9360
+    },
+    {
+      "epoch": 0.23425,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.0869140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4918,
+      "loss/crossentropy": 2.068470099568367,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19361184667795897,
+      "step": 9370
+    },
+    {
+      "epoch": 0.2345,
+      "grad_norm": 28.0,
+      "grad_norm_var": 6.534830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4265,
+      "loss/crossentropy": 2.125886672735214,
+      "loss/hidden": 3.401171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19355848152190447,
+      "step": 9380
+    },
+    {
+      "epoch": 0.23475,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.2770182291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3769,
+      "loss/crossentropy": 2.2057121500372885,
+      "loss/hidden": 3.28515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1843592157587409,
+      "step": 9390
+    },
+    {
+      "epoch": 0.235,
+      "grad_norm": 87.5,
+      "grad_norm_var": 202.43932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4444,
+      "loss/crossentropy": 2.17584248483181,
+      "loss/hidden": 3.4515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1986805137246847,
+      "step": 9400
+    },
+    {
+      "epoch": 0.23525,
+      "grad_norm": 33.0,
+      "grad_norm_var": 205.446875,
+      "learning_rate": 0.0001,
+      "loss": 7.5575,
+      "loss/crossentropy": 1.9989879056811333,
+      "loss/hidden": 3.484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21514312122017146,
+      "step": 9410
+    },
+    {
+      "epoch": 0.2355,
+      "grad_norm": 33.25,
+      "grad_norm_var": 3.4775390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3965,
+      "loss/crossentropy": 2.31248200237751,
+      "loss/hidden": 3.359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20192647576332093,
+      "step": 9420
+    },
+    {
+      "epoch": 0.23575,
+      "grad_norm": 34.0,
+      "grad_norm_var": 2.5723307291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.425,
+      "loss/crossentropy": 2.0246524304151534,
+      "loss/hidden": 3.43125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1936120893806219,
+      "step": 9430
+    },
+    {
+      "epoch": 0.236,
+      "grad_norm": 31.5,
+      "grad_norm_var": 3.8207682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3458,
+      "loss/crossentropy": 2.0714851915836334,
+      "loss/hidden": 3.315625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19567677434533834,
+      "step": 9440
+    },
+    {
+      "epoch": 0.23625,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.0452473958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3934,
+      "loss/crossentropy": 2.024892423301935,
+      "loss/hidden": 3.541796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19850265365093947,
+      "step": 9450
+    },
+    {
+      "epoch": 0.2365,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.16640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4192,
+      "loss/crossentropy": 2.146814212203026,
+      "loss/hidden": 3.43828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20048882197588683,
+      "step": 9460
+    },
+    {
+      "epoch": 0.23675,
+      "grad_norm": 27.875,
+      "grad_norm_var": 12.364322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3745,
+      "loss/crossentropy": 2.0004019677639007,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19019459020346402,
+      "step": 9470
+    },
+    {
+      "epoch": 0.237,
+      "grad_norm": 28.75,
+      "grad_norm_var": 12.885416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2858,
+      "loss/crossentropy": 2.0364832431077957,
+      "loss/hidden": 3.357421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18539349418133497,
+      "step": 9480
+    },
+    {
+      "epoch": 0.23725,
+      "grad_norm": 26.75,
+      "grad_norm_var": 4.488997395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2865,
+      "loss/crossentropy": 2.257227724790573,
+      "loss/hidden": 3.247265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1774066084995866,
+      "step": 9490
+    },
+    {
+      "epoch": 0.2375,
+      "grad_norm": 28.25,
+      "grad_norm_var": 5.15625,
+      "learning_rate": 0.0001,
+      "loss": 7.3742,
+      "loss/crossentropy": 2.192533364892006,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18531391881406306,
+      "step": 9500
+    },
+    {
+      "epoch": 0.23775,
+      "grad_norm": 28.5,
+      "grad_norm_var": 8.7025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4007,
+      "loss/crossentropy": 1.9954842567443847,
+      "loss/hidden": 3.512890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21250668447464705,
+      "step": 9510
+    },
+    {
+      "epoch": 0.238,
+      "grad_norm": 27.0,
+      "grad_norm_var": 3.2018229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3589,
+      "loss/crossentropy": 2.0501152604818342,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19774708338081837,
+      "step": 9520
+    },
+    {
+      "epoch": 0.23825,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.471875,
+      "learning_rate": 0.0001,
+      "loss": 7.3611,
+      "loss/crossentropy": 2.1446721121668815,
+      "loss/hidden": 3.40625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19521092902868986,
+      "step": 9530
+    },
+    {
+      "epoch": 0.2385,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.911393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4037,
+      "loss/crossentropy": 2.050874675065279,
+      "loss/hidden": 3.4078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18872325737029313,
+      "step": 9540
+    },
+    {
+      "epoch": 0.23875,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.69375,
+      "learning_rate": 0.0001,
+      "loss": 7.3851,
+      "loss/crossentropy": 2.092629846930504,
+      "loss/hidden": 3.3984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1858744696713984,
+      "step": 9550
+    },
+    {
+      "epoch": 0.239,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.6229166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.324,
+      "loss/crossentropy": 2.1389416724443437,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18115091007202863,
+      "step": 9560
+    },
+    {
+      "epoch": 0.23925,
+      "grad_norm": 28.375,
+      "grad_norm_var": 0.7893229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3684,
+      "loss/crossentropy": 2.1148220866918566,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18993774689733983,
+      "step": 9570
+    },
+    {
+      "epoch": 0.2395,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.0260416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4711,
+      "loss/crossentropy": 1.987958113849163,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19200538750737906,
+      "step": 9580
+    },
+    {
+      "epoch": 0.23975,
+      "grad_norm": 29.625,
+      "grad_norm_var": 0.6018229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3578,
+      "loss/crossentropy": 2.076607885956764,
+      "loss/hidden": 3.36640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20001101978123187,
+      "step": 9590
+    },
+    {
+      "epoch": 0.24,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.4124348958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4458,
+      "loss/crossentropy": 2.1496243715286254,
+      "loss/hidden": 3.357421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1889566643163562,
+      "step": 9600
+    },
+    {
+      "epoch": 0.24025,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.04140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4632,
+      "loss/crossentropy": 2.162689308822155,
+      "loss/hidden": 3.33828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19204493910074233,
+      "step": 9610
+    },
+    {
+      "epoch": 0.2405,
+      "grad_norm": 27.375,
+      "grad_norm_var": 1.5863932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3897,
+      "loss/crossentropy": 2.1323146484792233,
+      "loss/hidden": 3.455859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20718522872775794,
+      "step": 9620
+    },
+    {
+      "epoch": 0.24075,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.09375,
+      "learning_rate": 0.0001,
+      "loss": 7.2848,
+      "loss/crossentropy": 2.0804166465997698,
+      "loss/hidden": 3.417578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1966247998178005,
+      "step": 9630
+    },
+    {
+      "epoch": 0.241,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.3559895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4174,
+      "loss/crossentropy": 2.3036757931113243,
+      "loss/hidden": 3.341015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18416995517909526,
+      "step": 9640
+    },
+    {
+      "epoch": 0.24125,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.41015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2853,
+      "loss/crossentropy": 1.9936259984970093,
+      "loss/hidden": 3.428125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21977887880057095,
+      "step": 9650
+    },
+    {
+      "epoch": 0.2415,
+      "grad_norm": 32.5,
+      "grad_norm_var": 6.1228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4938,
+      "loss/crossentropy": 2.2000851720571517,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20301534831523896,
+      "step": 9660
+    },
+    {
+      "epoch": 0.24175,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.59765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5133,
+      "loss/crossentropy": 2.192278115451336,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19489070847630502,
+      "step": 9670
+    },
+    {
+      "epoch": 0.242,
+      "grad_norm": 28.875,
+      "grad_norm_var": 6.8978515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4247,
+      "loss/crossentropy": 2.194097451120615,
+      "loss/hidden": 3.43046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20746590523049235,
+      "step": 9680
+    },
+    {
+      "epoch": 0.24225,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.0587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3413,
+      "loss/crossentropy": 2.1902857303619383,
+      "loss/hidden": 3.187109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17439354099333287,
+      "step": 9690
+    },
+    {
+      "epoch": 0.2425,
+      "grad_norm": 41.5,
+      "grad_norm_var": 3.952541960104418e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5429,
+      "loss/crossentropy": 2.046304853260517,
+      "loss/hidden": 3.31953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18022917695343493,
+      "step": 9700
+    },
+    {
+      "epoch": 0.24275,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.952541960261809e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2729,
+      "loss/crossentropy": 2.082195009291172,
+      "loss/hidden": 3.5140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1983122780919075,
+      "step": 9710
+    },
+    {
+      "epoch": 0.243,
+      "grad_norm": 28.375,
+      "grad_norm_var": 2.6056640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2271,
+      "loss/crossentropy": 1.970278625190258,
+      "loss/hidden": 3.474609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19305091574788094,
+      "step": 9720
+    },
+    {
+      "epoch": 0.24325,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.0576524262052987e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3979,
+      "loss/crossentropy": 2.1140229746699335,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19698369763791562,
+      "step": 9730
+    },
+    {
+      "epoch": 0.2435,
+      "grad_norm": 29.0,
+      "grad_norm_var": 2.057652425978177e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2491,
+      "loss/crossentropy": 1.9204902969300748,
+      "loss/hidden": 3.40703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18290270324796437,
+      "step": 9740
+    },
+    {
+      "epoch": 0.24375,
+      "grad_norm": 49.25,
+      "grad_norm_var": 27.333072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4168,
+      "loss/crossentropy": 2.203968660533428,
+      "loss/hidden": 3.358203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19210707377642394,
+      "step": 9750
+    },
+    {
+      "epoch": 0.244,
+      "grad_norm": 27.75,
+      "grad_norm_var": 28.5837890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3667,
+      "loss/crossentropy": 2.0645239472389223,
+      "loss/hidden": 3.42578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1849998442456126,
+      "step": 9760
+    },
+    {
+      "epoch": 0.24425,
+      "grad_norm": 30.0,
+      "grad_norm_var": 7.441666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3958,
+      "loss/crossentropy": 2.0423281893134115,
+      "loss/hidden": 3.46640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19458430632948875,
+      "step": 9770
+    },
+    {
+      "epoch": 0.2445,
+      "grad_norm": 28.25,
+      "grad_norm_var": 25.725,
+      "learning_rate": 0.0001,
+      "loss": 7.371,
+      "loss/crossentropy": 2.2162967801094053,
+      "loss/hidden": 3.392578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19894264116883278,
+      "step": 9780
+    },
+    {
+      "epoch": 0.24475,
+      "grad_norm": 31.875,
+      "grad_norm_var": 25.701822916666668,
+      "learning_rate": 0.0001,
+      "loss": 7.4136,
+      "loss/crossentropy": 2.052091246843338,
+      "loss/hidden": 3.331640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19768061954528093,
+      "step": 9790
+    },
+    {
+      "epoch": 0.245,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.4247395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3752,
+      "loss/crossentropy": 2.0378331199288366,
+      "loss/hidden": 3.45703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1888222724199295,
+      "step": 9800
+    },
+    {
+      "epoch": 0.24525,
+      "grad_norm": 29.75,
+      "grad_norm_var": 17.073958333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.389,
+      "loss/crossentropy": 2.0574826121330263,
+      "loss/hidden": 3.313671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1817958688363433,
+      "step": 9810
+    },
+    {
+      "epoch": 0.2455,
+      "grad_norm": 28.125,
+      "grad_norm_var": 24.198372395833335,
+      "learning_rate": 0.0001,
+      "loss": 7.284,
+      "loss/crossentropy": 2.068347904086113,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1905125178396702,
+      "step": 9820
+    },
+    {
+      "epoch": 0.24575,
+      "grad_norm": 28.5,
+      "grad_norm_var": 9.8041015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3867,
+      "loss/crossentropy": 2.0113710410892964,
+      "loss/hidden": 3.318359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1825895557180047,
+      "step": 9830
+    },
+    {
+      "epoch": 0.246,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.16015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3496,
+      "loss/crossentropy": 2.0473890252411366,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18930337531492114,
+      "step": 9840
+    },
+    {
+      "epoch": 0.24625,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.75,
+      "learning_rate": 0.0001,
+      "loss": 7.4305,
+      "loss/crossentropy": 2.167546259611845,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18493952229619026,
+      "step": 9850
+    },
+    {
+      "epoch": 0.2465,
+      "grad_norm": 28.375,
+      "grad_norm_var": 2.7625,
+      "learning_rate": 0.0001,
+      "loss": 7.444,
+      "loss/crossentropy": 2.1046394810080526,
+      "loss/hidden": 3.324609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18062518630176783,
+      "step": 9860
+    },
+    {
+      "epoch": 0.24675,
+      "grad_norm": 28.75,
+      "grad_norm_var": 7819.056705729166,
+      "learning_rate": 0.0001,
+      "loss": 7.4016,
+      "loss/crossentropy": 2.0700221791863442,
+      "loss/hidden": 3.372265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.190355353243649,
+      "step": 9870
+    },
+    {
+      "epoch": 0.247,
+      "grad_norm": 29.0,
+      "grad_norm_var": 7843.72265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2875,
+      "loss/crossentropy": 2.057565826922655,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18850490506738424,
+      "step": 9880
+    },
+    {
+      "epoch": 0.24725,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.778059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4329,
+      "loss/crossentropy": 2.0141181223094464,
+      "loss/hidden": 3.446875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19769613686949014,
+      "step": 9890
+    },
+    {
+      "epoch": 0.2475,
+      "grad_norm": 30.375,
+      "grad_norm_var": 6.312239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.336,
+      "loss/crossentropy": 2.146551664918661,
+      "loss/hidden": 3.34296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18510441221296786,
+      "step": 9900
+    },
+    {
+      "epoch": 0.24775,
+      "grad_norm": 33.0,
+      "grad_norm_var": 6.36640625,
+      "learning_rate": 0.0001,
+      "loss": 7.306,
+      "loss/crossentropy": 2.2865438759326935,
+      "loss/hidden": 3.366796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19417236726731063,
+      "step": 9910
+    },
+    {
+      "epoch": 0.248,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.405989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3855,
+      "loss/crossentropy": 2.133404280245304,
+      "loss/hidden": 3.37578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19065556656569244,
+      "step": 9920
+    },
+    {
+      "epoch": 0.24825,
+      "grad_norm": 34.25,
+      "grad_norm_var": 3.1626528109486234e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4378,
+      "loss/crossentropy": 2.0854051023721696,
+      "loss/hidden": 3.459765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20351361632347106,
+      "step": 9930
+    },
+    {
+      "epoch": 0.2485,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.1626528099408717e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3826,
+      "loss/crossentropy": 2.087091060727835,
+      "loss/hidden": 3.39296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19672764679417015,
+      "step": 9940
+    },
+    {
+      "epoch": 0.24875,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.729622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4303,
+      "loss/crossentropy": 1.9529958970844745,
+      "loss/hidden": 3.461328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1957404987886548,
+      "step": 9950
+    },
+    {
+      "epoch": 0.249,
+      "grad_norm": 27.0,
+      "grad_norm_var": 3.515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2963,
+      "loss/crossentropy": 2.1291355013847353,
+      "loss/hidden": 3.278125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1832482174038887,
+      "step": 9960
+    },
+    {
+      "epoch": 0.24925,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.403580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3011,
+      "loss/crossentropy": 2.073649263381958,
+      "loss/hidden": 3.2859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17194083742797375,
+      "step": 9970
+    },
+    {
+      "epoch": 0.2495,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.7934895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3253,
+      "loss/crossentropy": 2.312130460143089,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19424791410565376,
+      "step": 9980
+    },
+    {
+      "epoch": 0.24975,
+      "grad_norm": 31.875,
+      "grad_norm_var": 6.4666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3395,
+      "loss/crossentropy": 1.9769040577113628,
+      "loss/hidden": 3.359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1865438589360565,
+      "step": 9990
+    },
+    {
+      "epoch": 0.25,
+      "grad_norm": 37.0,
+      "grad_norm_var": 5.488997395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3121,
+      "loss/crossentropy": 2.0656296610832214,
+      "loss/hidden": 3.418359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19510807991027831,
+      "step": 10000
+    },
+    {
+      "epoch": 0.25025,
+      "grad_norm": 30.5,
+      "grad_norm_var": 6.201822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3059,
+      "loss/crossentropy": 2.1552528128027917,
+      "loss/hidden": 3.338671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1871857862919569,
+      "step": 10010
+    },
+    {
+      "epoch": 0.2505,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.7697265625,
+      "learning_rate": 0.0001,
+      "loss": 7.402,
+      "loss/crossentropy": 2.1086191907525063,
+      "loss/hidden": 3.36953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20956437531858682,
+      "step": 10020
+    },
+    {
+      "epoch": 0.25075,
+      "grad_norm": 27.0,
+      "grad_norm_var": 79.871875,
+      "learning_rate": 0.0001,
+      "loss": 7.2906,
+      "loss/crossentropy": 2.24168366715312,
+      "loss/hidden": 3.333984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1914551755413413,
+      "step": 10030
+    },
+    {
+      "epoch": 0.251,
+      "grad_norm": 28.25,
+      "grad_norm_var": 3.5348307291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3933,
+      "loss/crossentropy": 1.9110309466719628,
+      "loss/hidden": 3.52421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1938932467252016,
+      "step": 10040
+    },
+    {
+      "epoch": 0.25125,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.318489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3066,
+      "loss/crossentropy": 2.2151729106903075,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1898823155090213,
+      "step": 10050
+    },
+    {
+      "epoch": 0.2515,
+      "grad_norm": 27.125,
+      "grad_norm_var": 5.114583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3306,
+      "loss/crossentropy": 1.9325201705098152,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1914183372631669,
+      "step": 10060
+    },
+    {
+      "epoch": 0.25175,
+      "grad_norm": 28.375,
+      "grad_norm_var": 5.568489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4302,
+      "loss/crossentropy": 2.053273378312588,
+      "loss/hidden": 3.48828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.193053549900651,
+      "step": 10070
+    },
+    {
+      "epoch": 0.252,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.5947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3679,
+      "loss/crossentropy": 2.1425619572401047,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19612068887799977,
+      "step": 10080
+    },
+    {
+      "epoch": 0.25225,
+      "grad_norm": 28.125,
+      "grad_norm_var": 1.2666666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4696,
+      "loss/crossentropy": 2.1615766674280166,
+      "loss/hidden": 3.437890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21496526505798103,
+      "step": 10090
+    },
+    {
+      "epoch": 0.2525,
+      "grad_norm": 31.375,
+      "grad_norm_var": 87.72389322916666,
+      "learning_rate": 0.0001,
+      "loss": 7.3933,
+      "loss/crossentropy": 2.093507520854473,
+      "loss/hidden": 3.306640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18447848707437514,
+      "step": 10100
+    },
+    {
+      "epoch": 0.25275,
+      "grad_norm": 29.125,
+      "grad_norm_var": 87.9962890625,
+      "learning_rate": 0.0001,
+      "loss": 7.42,
+      "loss/crossentropy": 2.2579648420214653,
+      "loss/hidden": 3.415625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19929521884769202,
+      "step": 10110
+    },
+    {
+      "epoch": 0.253,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.0921223958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4462,
+      "loss/crossentropy": 2.187203352898359,
+      "loss/hidden": 3.360546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18776683378964662,
+      "step": 10120
+    },
+    {
+      "epoch": 0.25325,
+      "grad_norm": 35.0,
+      "grad_norm_var": 37.43333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.493,
+      "loss/crossentropy": 2.156028524041176,
+      "loss/hidden": 3.41171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20926487501710653,
+      "step": 10130
+    },
+    {
+      "epoch": 0.2535,
+      "grad_norm": 30.625,
+      "grad_norm_var": 36.87180989583333,
+      "learning_rate": 0.0001,
+      "loss": 7.4227,
+      "loss/crossentropy": 1.9757653154432773,
+      "loss/hidden": 3.383984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1927978384308517,
+      "step": 10140
+    },
+    {
+      "epoch": 0.25375,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.1582682291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3403,
+      "loss/crossentropy": 2.1322492100298405,
+      "loss/hidden": 3.455859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1974259439855814,
+      "step": 10150
+    },
+    {
+      "epoch": 0.254,
+      "grad_norm": 28.125,
+      "grad_norm_var": 2.9259765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3083,
+      "loss/crossentropy": 2.0464363016188143,
+      "loss/hidden": 3.27734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17546537732705475,
+      "step": 10160
+    },
+    {
+      "epoch": 0.25425,
+      "grad_norm": 28.625,
+      "grad_norm_var": 3.7514973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3988,
+      "loss/crossentropy": 2.059445019811392,
+      "loss/hidden": 3.396484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19312301548197866,
+      "step": 10170
+    },
+    {
+      "epoch": 0.2545,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.36015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3178,
+      "loss/crossentropy": 2.0300123430788517,
+      "loss/hidden": 3.334375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18283235086128116,
+      "step": 10180
+    },
+    {
+      "epoch": 0.25475,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.7192057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4117,
+      "loss/crossentropy": 2.052568303793669,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17590052969753742,
+      "step": 10190
+    },
+    {
+      "epoch": 0.255,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.0660807291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4384,
+      "loss/crossentropy": 2.1124994076788424,
+      "loss/hidden": 3.4421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20564852859824895,
+      "step": 10200
+    },
+    {
+      "epoch": 0.25525,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.8223307291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4006,
+      "loss/crossentropy": 1.997454211115837,
+      "loss/hidden": 3.4078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.189224443025887,
+      "step": 10210
+    },
+    {
+      "epoch": 0.2555,
+      "grad_norm": 30.75,
+      "grad_norm_var": 12.555989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5754,
+      "loss/crossentropy": 2.095755438506603,
+      "loss/hidden": 3.470703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1905398152768612,
+      "step": 10220
+    },
+    {
+      "epoch": 0.25575,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.0666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4499,
+      "loss/crossentropy": 2.031408229470253,
+      "loss/hidden": 3.379296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1826389298774302,
+      "step": 10230
+    },
+    {
+      "epoch": 0.256,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.1973307291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4736,
+      "loss/crossentropy": 2.103855460882187,
+      "loss/hidden": 3.479296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21460786666721104,
+      "step": 10240
+    },
+    {
+      "epoch": 0.25625,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.6556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5078,
+      "loss/crossentropy": 2.0419384971261025,
+      "loss/hidden": 3.44609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1855954358354211,
+      "step": 10250
+    },
+    {
+      "epoch": 0.2565,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.025,
+      "learning_rate": 0.0001,
+      "loss": 7.3116,
+      "loss/crossentropy": 2.2164057731628417,
+      "loss/hidden": 3.336328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18515868298709393,
+      "step": 10260
+    },
+    {
+      "epoch": 0.25675,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.653059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2962,
+      "loss/crossentropy": 2.0967153370380402,
+      "loss/hidden": 3.5140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2144587781280279,
+      "step": 10270
+    },
+    {
+      "epoch": 0.257,
+      "grad_norm": 27.625,
+      "grad_norm_var": 2874.977018229167,
+      "learning_rate": 0.0001,
+      "loss": 7.3634,
+      "loss/crossentropy": 2.197367396950722,
+      "loss/hidden": 3.32421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1903674839064479,
+      "step": 10280
+    },
+    {
+      "epoch": 0.25725,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2830.1666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4461,
+      "loss/crossentropy": 1.9168385773897172,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18945675920695065,
+      "step": 10290
+    },
+    {
+      "epoch": 0.2575,
+      "grad_norm": 30.625,
+      "grad_norm_var": 10.3634765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4383,
+      "loss/crossentropy": 2.196033439040184,
+      "loss/hidden": 3.389453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20106703452765942,
+      "step": 10300
+    },
+    {
+      "epoch": 0.25775,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.41015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4,
+      "loss/crossentropy": 1.9811425000429153,
+      "loss/hidden": 3.2875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18086131382733583,
+      "step": 10310
+    },
+    {
+      "epoch": 0.258,
+      "grad_norm": 28.625,
+      "grad_norm_var": 1.3056640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3814,
+      "loss/crossentropy": 2.141839873790741,
+      "loss/hidden": 3.3515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19040859565138818,
+      "step": 10320
+    },
+    {
+      "epoch": 0.25825,
+      "grad_norm": 29.75,
+      "grad_norm_var": 0.9582682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3988,
+      "loss/crossentropy": 2.0412203505635262,
+      "loss/hidden": 3.471484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19841136487666516,
+      "step": 10330
+    },
+    {
+      "epoch": 0.2585,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.94765625,
+      "learning_rate": 0.0001,
+      "loss": 7.464,
+      "loss/crossentropy": 2.124895977973938,
+      "loss/hidden": 3.42890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19423805810511113,
+      "step": 10340
+    },
+    {
+      "epoch": 0.25875,
+      "grad_norm": 27.125,
+      "grad_norm_var": 20.603125,
+      "learning_rate": 0.0001,
+      "loss": 7.3612,
+      "loss/crossentropy": 2.271273523569107,
+      "loss/hidden": 3.341015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19086832217872143,
+      "step": 10350
+    },
+    {
+      "epoch": 0.259,
+      "grad_norm": 31.75,
+      "grad_norm_var": 41.87265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3645,
+      "loss/crossentropy": 2.148430307209492,
+      "loss/hidden": 3.33203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18722805399447678,
+      "step": 10360
+    },
+    {
+      "epoch": 0.25925,
+      "grad_norm": 31.125,
+      "grad_norm_var": 28.49140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2841,
+      "loss/crossentropy": 2.088160905241966,
+      "loss/hidden": 3.43359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19452540278434755,
+      "step": 10370
+    },
+    {
+      "epoch": 0.2595,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.8275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3269,
+      "loss/crossentropy": 2.037956405431032,
+      "loss/hidden": 3.324609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19043162725865842,
+      "step": 10380
+    },
+    {
+      "epoch": 0.25975,
+      "grad_norm": 28.25,
+      "grad_norm_var": 8.718489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2968,
+      "loss/crossentropy": 1.9823431193828582,
+      "loss/hidden": 3.36171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19301451742649078,
+      "step": 10390
+    },
+    {
+      "epoch": 0.26,
+      "grad_norm": 31.625,
+      "grad_norm_var": 8.5625,
+      "learning_rate": 0.0001,
+      "loss": 7.3314,
+      "loss/crossentropy": 2.0179347068071367,
+      "loss/hidden": 3.393359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18727120459079744,
+      "step": 10400
+    },
+    {
+      "epoch": 0.26025,
+      "grad_norm": 29.375,
+      "grad_norm_var": 7.3625,
+      "learning_rate": 0.0001,
+      "loss": 7.3878,
+      "loss/crossentropy": 2.0856245614588262,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18157497737556696,
+      "step": 10410
+    },
+    {
+      "epoch": 0.2605,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.0134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3355,
+      "loss/crossentropy": 2.0798820555210114,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1937016123905778,
+      "step": 10420
+    },
+    {
+      "epoch": 0.26075,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.5400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4527,
+      "loss/crossentropy": 2.0538913279771807,
+      "loss/hidden": 3.487890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19370690621435643,
+      "step": 10430
+    },
+    {
+      "epoch": 0.261,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.42890625,
+      "learning_rate": 0.0001,
+      "loss": 7.46,
+      "loss/crossentropy": 2.181585241854191,
+      "loss/hidden": 3.41953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1953628245741129,
+      "step": 10440
+    },
+    {
+      "epoch": 0.26125,
+      "grad_norm": 27.75,
+      "grad_norm_var": 3.783072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3934,
+      "loss/crossentropy": 2.183496044576168,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20456707030534743,
+      "step": 10450
+    },
+    {
+      "epoch": 0.2615,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.4379557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3211,
+      "loss/crossentropy": 2.132909268140793,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18731370605528355,
+      "step": 10460
+    },
+    {
+      "epoch": 0.26175,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.5791666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3587,
+      "loss/crossentropy": 2.0115842171013356,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18043918311595916,
+      "step": 10470
+    },
+    {
+      "epoch": 0.262,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.8268229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2962,
+      "loss/crossentropy": 2.1343379452824593,
+      "loss/hidden": 3.305859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18047874867916108,
+      "step": 10480
+    },
+    {
+      "epoch": 0.26225,
+      "grad_norm": 36.0,
+      "grad_norm_var": 8.967708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2903,
+      "loss/crossentropy": 1.8975451827049254,
+      "loss/hidden": 3.395703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17858757209032775,
+      "step": 10490
+    },
+    {
+      "epoch": 0.2625,
+      "grad_norm": 31.75,
+      "grad_norm_var": 10.077083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3074,
+      "loss/crossentropy": 2.0936397939920424,
+      "loss/hidden": 3.3828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19368096999824047,
+      "step": 10500
+    },
+    {
+      "epoch": 0.26275,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.020768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3675,
+      "loss/crossentropy": 2.06072843298316,
+      "loss/hidden": 3.30390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17674694433808327,
+      "step": 10510
+    },
+    {
+      "epoch": 0.263,
+      "grad_norm": 29.125,
+      "grad_norm_var": 3.4686848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3847,
+      "loss/crossentropy": 2.057940775156021,
+      "loss/hidden": 3.381640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18738706540316344,
+      "step": 10520
+    },
+    {
+      "epoch": 0.26325,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.225,
+      "learning_rate": 0.0001,
+      "loss": 7.3449,
+      "loss/crossentropy": 2.051366009563208,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18555977791547776,
+      "step": 10530
+    },
+    {
+      "epoch": 0.2635,
+      "grad_norm": 29.75,
+      "grad_norm_var": 32.99108072916667,
+      "learning_rate": 0.0001,
+      "loss": 7.3795,
+      "loss/crossentropy": 2.061921717226505,
+      "loss/hidden": 3.427734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1927286960184574,
+      "step": 10540
+    },
+    {
+      "epoch": 0.26375,
+      "grad_norm": 31.875,
+      "grad_norm_var": 34.16451822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.3305,
+      "loss/crossentropy": 2.051614002883434,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20516632981598376,
+      "step": 10550
+    },
+    {
+      "epoch": 0.264,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.2802083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3251,
+      "loss/crossentropy": 2.208609312772751,
+      "loss/hidden": 3.302734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1906878400593996,
+      "step": 10560
+    },
+    {
+      "epoch": 0.26425,
+      "grad_norm": 28.0,
+      "grad_norm_var": 2.8785807291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4584,
+      "loss/crossentropy": 2.2136432066559792,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1983118023723364,
+      "step": 10570
+    },
+    {
+      "epoch": 0.2645,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.562239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3017,
+      "loss/crossentropy": 2.2243916779756545,
+      "loss/hidden": 3.276953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19225324541330338,
+      "step": 10580
+    },
+    {
+      "epoch": 0.26475,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.47265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2623,
+      "loss/crossentropy": 2.085415804386139,
+      "loss/hidden": 3.308203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1769318761304021,
+      "step": 10590
+    },
+    {
+      "epoch": 0.265,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.8895182291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4543,
+      "loss/crossentropy": 2.0929081469774244,
+      "loss/hidden": 3.431640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21267817355692387,
+      "step": 10600
+    },
+    {
+      "epoch": 0.26525,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.505989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2136,
+      "loss/crossentropy": 1.885066507011652,
+      "loss/hidden": 3.34609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17078990247100592,
+      "step": 10610
+    },
+    {
+      "epoch": 0.2655,
+      "grad_norm": 27.25,
+      "grad_norm_var": 2.18515625,
+      "learning_rate": 0.0001,
+      "loss": 7.258,
+      "loss/crossentropy": 1.9193078093230724,
+      "loss/hidden": 3.366015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18096741195768118,
+      "step": 10620
+    },
+    {
+      "epoch": 0.26575,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.499739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.256,
+      "loss/crossentropy": 1.9643688909709454,
+      "loss/hidden": 3.317578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16975617725402117,
+      "step": 10630
+    },
+    {
+      "epoch": 0.266,
+      "grad_norm": 41.25,
+      "grad_norm_var": 2.3308942592018744e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3436,
+      "loss/crossentropy": 1.8120718583464623,
+      "loss/hidden": 3.403515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17283489797264337,
+      "step": 10640
+    },
+    {
+      "epoch": 0.26625,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.3308942571726003e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.398,
+      "loss/crossentropy": 2.2462258487939835,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20116339288651944,
+      "step": 10650
+    },
+    {
+      "epoch": 0.2665,
+      "grad_norm": 32.0,
+      "grad_norm_var": 21.258268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2994,
+      "loss/crossentropy": 2.044428373128176,
+      "loss/hidden": 3.42265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1989501324482262,
+      "step": 10660
+    },
+    {
+      "epoch": 0.26675,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.8681640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3881,
+      "loss/crossentropy": 1.9573393151164056,
+      "loss/hidden": 3.27421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17021252401173115,
+      "step": 10670
+    },
+    {
+      "epoch": 0.267,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.6197265625,
+      "learning_rate": 0.0001,
+      "loss": 7.449,
+      "loss/crossentropy": 2.1520077705383303,
+      "loss/hidden": 3.498046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20022546350955964,
+      "step": 10680
+    },
+    {
+      "epoch": 0.26725,
+      "grad_norm": 30.25,
+      "grad_norm_var": 69.97473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4428,
+      "loss/crossentropy": 2.1147203534841537,
+      "loss/hidden": 3.366015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18886133264750243,
+      "step": 10690
+    },
+    {
+      "epoch": 0.2675,
+      "grad_norm": 33.25,
+      "grad_norm_var": 27.978125,
+      "learning_rate": 0.0001,
+      "loss": 7.3953,
+      "loss/crossentropy": 2.1656775265932082,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18777367267757655,
+      "step": 10700
+    },
+    {
+      "epoch": 0.26775,
+      "grad_norm": 30.75,
+      "grad_norm_var": 39.307291666666664,
+      "learning_rate": 0.0001,
+      "loss": 7.2137,
+      "loss/crossentropy": 2.2314133137464522,
+      "loss/hidden": 3.280859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18172509353607894,
+      "step": 10710
+    },
+    {
+      "epoch": 0.268,
+      "grad_norm": 30.5,
+      "grad_norm_var": 43.19055989583333,
+      "learning_rate": 0.0001,
+      "loss": 7.3973,
+      "loss/crossentropy": 2.0373408157378434,
+      "loss/hidden": 3.500390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19881470818072558,
+      "step": 10720
+    },
+    {
+      "epoch": 0.26825,
+      "grad_norm": 28.125,
+      "grad_norm_var": 7.35625,
+      "learning_rate": 0.0001,
+      "loss": 7.3064,
+      "loss/crossentropy": 2.065886814892292,
+      "loss/hidden": 3.328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19001882169395684,
+      "step": 10730
+    },
+    {
+      "epoch": 0.2685,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.437239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4337,
+      "loss/crossentropy": 2.1057795181870462,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19459401965141296,
+      "step": 10740
+    },
+    {
+      "epoch": 0.26875,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.24765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4157,
+      "loss/crossentropy": 1.7320681288838387,
+      "loss/hidden": 3.533984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20864265877753496,
+      "step": 10750
+    },
+    {
+      "epoch": 0.269,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.9098307291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3267,
+      "loss/crossentropy": 2.1352997794747353,
+      "loss/hidden": 3.398828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19193184426985682,
+      "step": 10760
+    },
+    {
+      "epoch": 0.26925,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.0444333193999944e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3861,
+      "loss/crossentropy": 2.070178285241127,
+      "loss/hidden": 3.4390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19077087007462978,
+      "step": 10770
+    },
+    {
+      "epoch": 0.2695,
+      "grad_norm": 28.5,
+      "grad_norm_var": 23.91640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4494,
+      "loss/crossentropy": 2.147341425716877,
+      "loss/hidden": 3.341015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18359180726110935,
+      "step": 10780
+    },
+    {
+      "epoch": 0.26975,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.5843098958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3566,
+      "loss/crossentropy": 2.0286200530827045,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1995188482105732,
+      "step": 10790
+    },
+    {
+      "epoch": 0.27,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.8247395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3752,
+      "loss/crossentropy": 2.1403896272182465,
+      "loss/hidden": 3.408203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19066570214927198,
+      "step": 10800
+    },
+    {
+      "epoch": 0.27025,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.2462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3297,
+      "loss/crossentropy": 2.109960842132568,
+      "loss/hidden": 3.35234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19408894181251526,
+      "step": 10810
+    },
+    {
+      "epoch": 0.2705,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.125,
+      "learning_rate": 0.0001,
+      "loss": 7.4115,
+      "loss/crossentropy": 2.170359855145216,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1864105872809887,
+      "step": 10820
+    },
+    {
+      "epoch": 0.27075,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.5025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3568,
+      "loss/crossentropy": 2.0696767389774324,
+      "loss/hidden": 3.32109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18454026989638805,
+      "step": 10830
+    },
+    {
+      "epoch": 0.271,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.187239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3986,
+      "loss/crossentropy": 2.1671934336423875,
+      "loss/hidden": 3.377734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19040703494101763,
+      "step": 10840
+    },
+    {
+      "epoch": 0.27125,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.470833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3105,
+      "loss/crossentropy": 1.9061052799224854,
+      "loss/hidden": 3.3734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1798303933814168,
+      "step": 10850
+    },
+    {
+      "epoch": 0.2715,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.2087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3692,
+      "loss/crossentropy": 2.1218462653458117,
+      "loss/hidden": 3.432421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20523111913353204,
+      "step": 10860
+    },
+    {
+      "epoch": 0.27175,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.544205729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3206,
+      "loss/crossentropy": 2.006873355805874,
+      "loss/hidden": 3.4734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2144127245992422,
+      "step": 10870
+    },
+    {
+      "epoch": 0.272,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.5832967245463552e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4255,
+      "loss/crossentropy": 2.1331238821148872,
+      "loss/hidden": 3.366015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19302980024367572,
+      "step": 10880
+    },
+    {
+      "epoch": 0.27225,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.5832967244414976e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3477,
+      "loss/crossentropy": 2.041455736756325,
+      "loss/hidden": 3.4515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.230126572214067,
+      "step": 10890
+    },
+    {
+      "epoch": 0.2725,
+      "grad_norm": 32.25,
+      "grad_norm_var": 16.21015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3668,
+      "loss/crossentropy": 2.11685880869627,
+      "loss/hidden": 3.32890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18238812778145075,
+      "step": 10900
+    },
+    {
+      "epoch": 0.27275,
+      "grad_norm": 29.625,
+      "grad_norm_var": 15.566666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.349,
+      "loss/crossentropy": 2.1108590021729468,
+      "loss/hidden": 3.448828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2031965653412044,
+      "step": 10910
+    },
+    {
+      "epoch": 0.273,
+      "grad_norm": 30.5,
+      "grad_norm_var": 6.133072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3685,
+      "loss/crossentropy": 1.9768708415329457,
+      "loss/hidden": 3.4,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1786771345883608,
+      "step": 10920
+    },
+    {
+      "epoch": 0.27325,
+      "grad_norm": 28.75,
+      "grad_norm_var": 6.448372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3505,
+      "loss/crossentropy": 1.9710132874548436,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18313440950587392,
+      "step": 10930
+    },
+    {
+      "epoch": 0.2735,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.63515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4147,
+      "loss/crossentropy": 2.1898112446069717,
+      "loss/hidden": 3.3546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1952790988609195,
+      "step": 10940
+    },
+    {
+      "epoch": 0.27375,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.8978515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3679,
+      "loss/crossentropy": 2.135420022904873,
+      "loss/hidden": 3.43203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21568923965096473,
+      "step": 10950
+    },
+    {
+      "epoch": 0.274,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.5184895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2611,
+      "loss/crossentropy": 1.9667419284582137,
+      "loss/hidden": 3.42265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1834744794294238,
+      "step": 10960
+    },
+    {
+      "epoch": 0.27425,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.3561848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4273,
+      "loss/crossentropy": 2.074997512996197,
+      "loss/hidden": 3.356640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18813611660152674,
+      "step": 10970
+    },
+    {
+      "epoch": 0.2745,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.325455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3688,
+      "loss/crossentropy": 2.234239089488983,
+      "loss/hidden": 3.254296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18571207812055945,
+      "step": 10980
+    },
+    {
+      "epoch": 0.27475,
+      "grad_norm": 33.25,
+      "grad_norm_var": 5.548958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4528,
+      "loss/crossentropy": 2.1355573579669,
+      "loss/hidden": 3.489453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20484505780041218,
+      "step": 10990
+    },
+    {
+      "epoch": 0.275,
+      "grad_norm": 26.625,
+      "grad_norm_var": 5.2759765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4261,
+      "loss/crossentropy": 2.2509318992495535,
+      "loss/hidden": 3.3265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19250140003859997,
+      "step": 11000
+    },
+    {
+      "epoch": 0.27525,
+      "grad_norm": 31.0,
+      "grad_norm_var": 5.387239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3389,
+      "loss/crossentropy": 2.114951176941395,
+      "loss/hidden": 3.3640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.198347245156765,
+      "step": 11010
+    },
+    {
+      "epoch": 0.2755,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.5468098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3199,
+      "loss/crossentropy": 2.110514160990715,
+      "loss/hidden": 3.381640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1900305664166808,
+      "step": 11020
+    },
+    {
+      "epoch": 0.27575,
+      "grad_norm": 27.25,
+      "grad_norm_var": 2.003059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3304,
+      "loss/crossentropy": 2.201597589254379,
+      "loss/hidden": 3.393359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20035982932895421,
+      "step": 11030
+    },
+    {
+      "epoch": 0.276,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.0139973958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4256,
+      "loss/crossentropy": 2.1428385630249975,
+      "loss/hidden": 3.287890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18116684164851904,
+      "step": 11040
+    },
+    {
+      "epoch": 0.27625,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.4671223958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2733,
+      "loss/crossentropy": 2.043394061923027,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17936107851564884,
+      "step": 11050
+    },
+    {
+      "epoch": 0.2765,
+      "grad_norm": 32.75,
+      "grad_norm_var": 1.3830729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4904,
+      "loss/crossentropy": 2.1975908786058427,
+      "loss/hidden": 3.392578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21010352578014135,
+      "step": 11060
+    },
+    {
+      "epoch": 0.27675,
+      "grad_norm": 30.0,
+      "grad_norm_var": 4.671875,
+      "learning_rate": 0.0001,
+      "loss": 7.3718,
+      "loss/crossentropy": 2.1417069509625435,
+      "loss/hidden": 3.35390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17585361283272505,
+      "step": 11070
+    },
+    {
+      "epoch": 0.277,
+      "grad_norm": 29.375,
+      "grad_norm_var": 4.309375,
+      "learning_rate": 0.0001,
+      "loss": 7.2855,
+      "loss/crossentropy": 2.041763362288475,
+      "loss/hidden": 3.221875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17130352668464183,
+      "step": 11080
+    },
+    {
+      "epoch": 0.27725,
+      "grad_norm": 28.375,
+      "grad_norm_var": 11.15390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3131,
+      "loss/crossentropy": 2.1522015750408174,
+      "loss/hidden": 3.3375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18296321779489516,
+      "step": 11090
+    },
+    {
+      "epoch": 0.2775,
+      "grad_norm": 30.375,
+      "grad_norm_var": 11.266080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3499,
+      "loss/crossentropy": 2.1265663146972655,
+      "loss/hidden": 3.318359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18307223506271839,
+      "step": 11100
+    },
+    {
+      "epoch": 0.27775,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.6994140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3764,
+      "loss/crossentropy": 2.126886320114136,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1963003095239401,
+      "step": 11110
+    },
+    {
+      "epoch": 0.278,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.9780598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3686,
+      "loss/crossentropy": 2.209390402585268,
+      "loss/hidden": 3.284375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18762579131871462,
+      "step": 11120
+    },
+    {
+      "epoch": 0.27825,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.3806640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4086,
+      "loss/crossentropy": 2.0636716064065697,
+      "loss/hidden": 3.4109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18702280572615565,
+      "step": 11130
+    },
+    {
+      "epoch": 0.2785,
+      "grad_norm": 30.0,
+      "grad_norm_var": 11.084375,
+      "learning_rate": 0.0001,
+      "loss": 7.3611,
+      "loss/crossentropy": 1.9728175386786462,
+      "loss/hidden": 3.371875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18249935954809188,
+      "step": 11140
+    },
+    {
+      "epoch": 0.27875,
+      "grad_norm": 30.75,
+      "grad_norm_var": 13.183333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2901,
+      "loss/crossentropy": 1.9563469380140304,
+      "loss/hidden": 3.32421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1720572842285037,
+      "step": 11150
+    },
+    {
+      "epoch": 0.279,
+      "grad_norm": 40.75,
+      "grad_norm_var": 11.42890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2633,
+      "loss/crossentropy": 2.1441919445991515,
+      "loss/hidden": 3.30703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1793568328022957,
+      "step": 11160
+    },
+    {
+      "epoch": 0.27925,
+      "grad_norm": 31.125,
+      "grad_norm_var": 9.245833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3546,
+      "loss/crossentropy": 2.0893258482217787,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18886512089520693,
+      "step": 11170
+    },
+    {
+      "epoch": 0.2795,
+      "grad_norm": 29.25,
+      "grad_norm_var": 15.84140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3265,
+      "loss/crossentropy": 2.028592649102211,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1916877321898937,
+      "step": 11180
+    },
+    {
+      "epoch": 0.27975,
+      "grad_norm": 29.0,
+      "grad_norm_var": 5.451822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4362,
+      "loss/crossentropy": 2.1735318168997764,
+      "loss/hidden": 3.325,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18261966463178397,
+      "step": 11190
+    },
+    {
+      "epoch": 0.28,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.405143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3648,
+      "loss/crossentropy": 2.129372274875641,
+      "loss/hidden": 3.434375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19846085608005523,
+      "step": 11200
+    },
+    {
+      "epoch": 0.28025,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.4462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4302,
+      "loss/crossentropy": 2.1702652648091316,
+      "loss/hidden": 3.466015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20027328319847584,
+      "step": 11210
+    },
+    {
+      "epoch": 0.2805,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.58515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3575,
+      "loss/crossentropy": 2.134139196574688,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1903523735702038,
+      "step": 11220
+    },
+    {
+      "epoch": 0.28075,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.1093098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.43,
+      "loss/crossentropy": 2.026133489608765,
+      "loss/hidden": 3.45390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.201400394923985,
+      "step": 11230
+    },
+    {
+      "epoch": 0.281,
+      "grad_norm": 31.25,
+      "grad_norm_var": 12.038997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4162,
+      "loss/crossentropy": 2.1442901253700257,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20070651173591614,
+      "step": 11240
+    },
+    {
+      "epoch": 0.28125,
+      "grad_norm": 28.875,
+      "grad_norm_var": 162.22962239583333,
+      "learning_rate": 0.0001,
+      "loss": 7.4058,
+      "loss/crossentropy": 2.1855412632226945,
+      "loss/hidden": 3.4109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1976642705500126,
+      "step": 11250
+    },
+    {
+      "epoch": 0.2815,
+      "grad_norm": 31.75,
+      "grad_norm_var": 10.683072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3535,
+      "loss/crossentropy": 2.030480499565601,
+      "loss/hidden": 3.40078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20611500162631274,
+      "step": 11260
+    },
+    {
+      "epoch": 0.28175,
+      "grad_norm": 36.75,
+      "grad_norm_var": 11.08515625,
+      "learning_rate": 0.0001,
+      "loss": 7.287,
+      "loss/crossentropy": 2.01583767645061,
+      "loss/hidden": 3.312109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18007306316867472,
+      "step": 11270
+    },
+    {
+      "epoch": 0.282,
+      "grad_norm": 30.625,
+      "grad_norm_var": 7.605989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3012,
+      "loss/crossentropy": 2.2034949243068693,
+      "loss/hidden": 3.327734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18539215214550495,
+      "step": 11280
+    },
+    {
+      "epoch": 0.28225,
+      "grad_norm": 29.125,
+      "grad_norm_var": 4.1791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3566,
+      "loss/crossentropy": 2.0296434491872786,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18707384672015906,
+      "step": 11290
+    },
+    {
+      "epoch": 0.2825,
+      "grad_norm": 30.375,
+      "grad_norm_var": 7.01015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2793,
+      "loss/crossentropy": 2.0983324527740477,
+      "loss/hidden": 3.327734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17865268159657716,
+      "step": 11300
+    },
+    {
+      "epoch": 0.28275,
+      "grad_norm": 34.0,
+      "grad_norm_var": 7.029622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3716,
+      "loss/crossentropy": 2.127777361869812,
+      "loss/hidden": 3.476953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20420280396938323,
+      "step": 11310
+    },
+    {
+      "epoch": 0.283,
+      "grad_norm": 32.5,
+      "grad_norm_var": 38.66712239583333,
+      "learning_rate": 0.0001,
+      "loss": 7.3326,
+      "loss/crossentropy": 2.1566628091037274,
+      "loss/hidden": 3.28828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.180232659727335,
+      "step": 11320
+    },
+    {
+      "epoch": 0.28325,
+      "grad_norm": 30.5,
+      "grad_norm_var": 44.13743489583333,
+      "learning_rate": 0.0001,
+      "loss": 7.3567,
+      "loss/crossentropy": 2.091618612408638,
+      "loss/hidden": 3.330078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18182430015876888,
+      "step": 11330
+    },
+    {
+      "epoch": 0.2835,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.887434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3312,
+      "loss/crossentropy": 2.1611782908439636,
+      "loss/hidden": 3.369140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1828035417944193,
+      "step": 11340
+    },
+    {
+      "epoch": 0.28375,
+      "grad_norm": 34.25,
+      "grad_norm_var": 6.423893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.1759,
+      "loss/crossentropy": 1.9736236594617367,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18899882938712836,
+      "step": 11350
+    },
+    {
+      "epoch": 0.284,
+      "grad_norm": 30.0,
+      "grad_norm_var": 9.488541666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.1771,
+      "loss/crossentropy": 2.0265045419335364,
+      "loss/hidden": 3.273828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1712389207445085,
+      "step": 11360
+    },
+    {
+      "epoch": 0.28425,
+      "grad_norm": 33.75,
+      "grad_norm_var": 12.092122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.361,
+      "loss/crossentropy": 2.2598242223262788,
+      "loss/hidden": 3.3265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19073396287858485,
+      "step": 11370
+    },
+    {
+      "epoch": 0.2845,
+      "grad_norm": 29.0,
+      "grad_norm_var": 8.455989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3234,
+      "loss/crossentropy": 2.144850969314575,
+      "loss/hidden": 3.321484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18078193911351265,
+      "step": 11380
+    },
+    {
+      "epoch": 0.28475,
+      "grad_norm": 29.25,
+      "grad_norm_var": 1.7212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3,
+      "loss/crossentropy": 2.1646609872579576,
+      "loss/hidden": 3.3140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18855173885822296,
+      "step": 11390
+    },
+    {
+      "epoch": 0.285,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.5837890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2976,
+      "loss/crossentropy": 2.120090515911579,
+      "loss/hidden": 3.308203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17507072482258082,
+      "step": 11400
+    },
+    {
+      "epoch": 0.28525,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.0785807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4018,
+      "loss/crossentropy": 2.0800730645656587,
+      "loss/hidden": 3.43046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20872990731149912,
+      "step": 11410
+    },
+    {
+      "epoch": 0.2855,
+      "grad_norm": 30.25,
+      "grad_norm_var": 13.87890625,
+      "learning_rate": 0.0001,
+      "loss": 7.1549,
+      "loss/crossentropy": 2.1735212251544,
+      "loss/hidden": 3.25625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18148906547576188,
+      "step": 11420
+    },
+    {
+      "epoch": 0.28575,
+      "grad_norm": 30.375,
+      "grad_norm_var": 15.132747395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.206,
+      "loss/crossentropy": 1.8314741916954518,
+      "loss/hidden": 3.272265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17273464631289243,
+      "step": 11430
+    },
+    {
+      "epoch": 0.286,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.2330729166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3502,
+      "loss/crossentropy": 1.9636109337210654,
+      "loss/hidden": 3.329296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16999556813389063,
+      "step": 11440
+    },
+    {
+      "epoch": 0.28625,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.8009765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2893,
+      "loss/crossentropy": 1.9146189287304878,
+      "loss/hidden": 3.425390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1754764079116285,
+      "step": 11450
+    },
+    {
+      "epoch": 0.2865,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.9197265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3233,
+      "loss/crossentropy": 2.0301330149173737,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1742406915873289,
+      "step": 11460
+    },
+    {
+      "epoch": 0.28675,
+      "grad_norm": 34.25,
+      "grad_norm_var": 3.8009765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4231,
+      "loss/crossentropy": 2.1827415734529496,
+      "loss/hidden": 3.33515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18332260251045226,
+      "step": 11470
+    },
+    {
+      "epoch": 0.287,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.479622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3842,
+      "loss/crossentropy": 1.9815428338944912,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17839936017990113,
+      "step": 11480
+    },
+    {
+      "epoch": 0.28725,
+      "grad_norm": 30.125,
+      "grad_norm_var": 12.9228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3888,
+      "loss/crossentropy": 2.10543007850647,
+      "loss/hidden": 3.29765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18028822876513004,
+      "step": 11490
+    },
+    {
+      "epoch": 0.2875,
+      "grad_norm": 30.5,
+      "grad_norm_var": 0.9973307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2278,
+      "loss/crossentropy": 1.9690312303602695,
+      "loss/hidden": 3.35390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19321996718645096,
+      "step": 11500
+    },
+    {
+      "epoch": 0.28775,
+      "grad_norm": 29.5,
+      "grad_norm_var": 7.462239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4485,
+      "loss/crossentropy": 2.0997009545564653,
+      "loss/hidden": 3.371875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1947010463103652,
+      "step": 11510
+    },
+    {
+      "epoch": 0.288,
+      "grad_norm": 30.625,
+      "grad_norm_var": 6.814583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4014,
+      "loss/crossentropy": 2.041436542570591,
+      "loss/hidden": 3.371484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17614111984148623,
+      "step": 11520
+    },
+    {
+      "epoch": 0.28825,
+      "grad_norm": 29.625,
+      "grad_norm_var": 1.7247395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3231,
+      "loss/crossentropy": 2.056125995516777,
+      "loss/hidden": 3.475,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20333750136196613,
+      "step": 11530
+    },
+    {
+      "epoch": 0.2885,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.0014973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2735,
+      "loss/crossentropy": 1.9806349158287049,
+      "loss/hidden": 3.306640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17432715664617718,
+      "step": 11540
+    },
+    {
+      "epoch": 0.28875,
+      "grad_norm": 32.75,
+      "grad_norm_var": 1.6973307291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3842,
+      "loss/crossentropy": 2.0468244731426237,
+      "loss/hidden": 3.506640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19939990881830455,
+      "step": 11550
+    },
+    {
+      "epoch": 0.289,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.546875,
+      "learning_rate": 0.0001,
+      "loss": 7.4212,
+      "loss/crossentropy": 2.1170118719339373,
+      "loss/hidden": 3.452734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1938545262441039,
+      "step": 11560
+    },
+    {
+      "epoch": 0.28925,
+      "grad_norm": 29.0,
+      "grad_norm_var": 6.987239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3068,
+      "loss/crossentropy": 1.9378852248191833,
+      "loss/hidden": 3.392578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18525508288294076,
+      "step": 11570
+    },
+    {
+      "epoch": 0.2895,
+      "grad_norm": 32.25,
+      "grad_norm_var": 7.7650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3939,
+      "loss/crossentropy": 2.094913274049759,
+      "loss/hidden": 3.328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19857972972095012,
+      "step": 11580
+    },
+    {
+      "epoch": 0.28975,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.3934895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3634,
+      "loss/crossentropy": 2.112175312638283,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1923027027398348,
+      "step": 11590
+    },
+    {
+      "epoch": 0.29,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.8676432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4788,
+      "loss/crossentropy": 2.347763030230999,
+      "loss/hidden": 3.379296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20385736282914876,
+      "step": 11600
+    },
+    {
+      "epoch": 0.29025,
+      "grad_norm": 36.0,
+      "grad_norm_var": 2.1060661210680481e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.482,
+      "loss/crossentropy": 2.1438632771372794,
+      "loss/hidden": 3.375390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1947392811998725,
+      "step": 11610
+    },
+    {
+      "epoch": 0.2905,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.1060661214006216e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2853,
+      "loss/crossentropy": 2.080791215598583,
+      "loss/hidden": 3.2875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19497475158423186,
+      "step": 11620
+    },
+    {
+      "epoch": 0.29075,
+      "grad_norm": 30.0,
+      "grad_norm_var": 1.5931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2971,
+      "loss/crossentropy": 2.16863374710083,
+      "loss/hidden": 3.30390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18860027231276036,
+      "step": 11630
+    },
+    {
+      "epoch": 0.291,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.5994140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3835,
+      "loss/crossentropy": 2.061967647075653,
+      "loss/hidden": 3.366015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18620339650660753,
+      "step": 11640
+    },
+    {
+      "epoch": 0.29125,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.690207280948122e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3738,
+      "loss/crossentropy": 2.0998649850487707,
+      "loss/hidden": 3.279296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18567133340984582,
+      "step": 11650
+    },
+    {
+      "epoch": 0.2915,
+      "grad_norm": 29.0,
+      "grad_norm_var": 29.6228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2405,
+      "loss/crossentropy": 2.1259921073913572,
+      "loss/hidden": 3.326953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18609032463282346,
+      "step": 11660
+    },
+    {
+      "epoch": 0.29175,
+      "grad_norm": 31.0,
+      "grad_norm_var": 14.421809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4326,
+      "loss/crossentropy": 2.066218351200223,
+      "loss/hidden": 3.348828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1920524787157774,
+      "step": 11670
+    },
+    {
+      "epoch": 0.292,
+      "grad_norm": 27.875,
+      "grad_norm_var": 49.305989583333336,
+      "learning_rate": 0.0001,
+      "loss": 7.2881,
+      "loss/crossentropy": 2.07202163413167,
+      "loss/hidden": 3.35,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18048244724050164,
+      "step": 11680
+    },
+    {
+      "epoch": 0.29225,
+      "grad_norm": 31.875,
+      "grad_norm_var": 4.995247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2687,
+      "loss/crossentropy": 2.045185898244381,
+      "loss/hidden": 3.275,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17395208198577167,
+      "step": 11690
+    },
+    {
+      "epoch": 0.2925,
+      "grad_norm": 30.0,
+      "grad_norm_var": 5.696875,
+      "learning_rate": 0.0001,
+      "loss": 7.3259,
+      "loss/crossentropy": 2.3408665537834166,
+      "loss/hidden": 3.345703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20136666856706142,
+      "step": 11700
+    },
+    {
+      "epoch": 0.29275,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.7421223958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2137,
+      "loss/crossentropy": 1.9074330985546113,
+      "loss/hidden": 3.349609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1704694928601384,
+      "step": 11710
+    },
+    {
+      "epoch": 0.293,
+      "grad_norm": 31.5,
+      "grad_norm_var": 5.803125,
+      "learning_rate": 0.0001,
+      "loss": 7.2817,
+      "loss/crossentropy": 2.290114316344261,
+      "loss/hidden": 3.36484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20061923637986184,
+      "step": 11720
+    },
+    {
+      "epoch": 0.29325,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.4936848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4136,
+      "loss/crossentropy": 2.2652835667133333,
+      "loss/hidden": 3.29375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18269798178225755,
+      "step": 11730
+    },
+    {
+      "epoch": 0.2935,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.06640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3219,
+      "loss/crossentropy": 2.047122722864151,
+      "loss/hidden": 3.361328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18665554728358985,
+      "step": 11740
+    },
+    {
+      "epoch": 0.29375,
+      "grad_norm": 29.875,
+      "grad_norm_var": 11.484309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4668,
+      "loss/crossentropy": 2.0893666088581084,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18700389545410873,
+      "step": 11750
+    },
+    {
+      "epoch": 0.294,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.412239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3389,
+      "loss/crossentropy": 2.0646077781915664,
+      "loss/hidden": 3.3890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1896628426387906,
+      "step": 11760
+    },
+    {
+      "epoch": 0.29425,
+      "grad_norm": 36.75,
+      "grad_norm_var": 15.843489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3537,
+      "loss/crossentropy": 2.14951953291893,
+      "loss/hidden": 3.420703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19772466979920864,
+      "step": 11770
+    },
+    {
+      "epoch": 0.2945,
+      "grad_norm": 34.0,
+      "grad_norm_var": 15.808333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2266,
+      "loss/crossentropy": 1.9084580048918725,
+      "loss/hidden": 3.31953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17457483559846879,
+      "step": 11780
+    },
+    {
+      "epoch": 0.29475,
+      "grad_norm": 30.5,
+      "grad_norm_var": 17.133268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4198,
+      "loss/crossentropy": 2.1059816129505635,
+      "loss/hidden": 3.2609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19060559011995792,
+      "step": 11790
+    },
+    {
+      "epoch": 0.295,
+      "grad_norm": 28.875,
+      "grad_norm_var": 15.191666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3353,
+      "loss/crossentropy": 2.0854270696640014,
+      "loss/hidden": 3.340625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18200992513448,
+      "step": 11800
+    },
+    {
+      "epoch": 0.29525,
+      "grad_norm": 33.0,
+      "grad_norm_var": 5.77890625,
+      "learning_rate": 0.0001,
+      "loss": 7.338,
+      "loss/crossentropy": 2.0409404814243315,
+      "loss/hidden": 3.3640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18492084443569184,
+      "step": 11810
+    },
+    {
+      "epoch": 0.2955,
+      "grad_norm": 28.25,
+      "grad_norm_var": 4.9556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2394,
+      "loss/crossentropy": 2.1242421194911003,
+      "loss/hidden": 3.318359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17592983674257995,
+      "step": 11820
+    },
+    {
+      "epoch": 0.29575,
+      "grad_norm": 32.75,
+      "grad_norm_var": 12.3994140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3172,
+      "loss/crossentropy": 2.135970714688301,
+      "loss/hidden": 3.328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1905512258410454,
+      "step": 11830
+    },
+    {
+      "epoch": 0.296,
+      "grad_norm": 31.125,
+      "grad_norm_var": 13.479166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3928,
+      "loss/crossentropy": 2.0788447827100756,
+      "loss/hidden": 3.35859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18583104945719242,
+      "step": 11840
+    },
+    {
+      "epoch": 0.29625,
+      "grad_norm": 29.375,
+      "grad_norm_var": 5.16640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2575,
+      "loss/crossentropy": 2.0609687596559523,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1885878125205636,
+      "step": 11850
+    },
+    {
+      "epoch": 0.2965,
+      "grad_norm": 33.5,
+      "grad_norm_var": 4.534309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2397,
+      "loss/crossentropy": 2.113059785962105,
+      "loss/hidden": 3.347265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19279801957309245,
+      "step": 11860
+    },
+    {
+      "epoch": 0.29675,
+      "grad_norm": 28.0,
+      "grad_norm_var": 2.833268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4813,
+      "loss/crossentropy": 2.2016214139759542,
+      "loss/hidden": 3.325,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19246475584805012,
+      "step": 11870
+    },
+    {
+      "epoch": 0.297,
+      "grad_norm": 35.5,
+      "grad_norm_var": 5.829622395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4107,
+      "loss/crossentropy": 2.1937885224819182,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20546972285956144,
+      "step": 11880
+    },
+    {
+      "epoch": 0.29725,
+      "grad_norm": 32.25,
+      "grad_norm_var": 5.101041666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4133,
+      "loss/crossentropy": 2.112873890995979,
+      "loss/hidden": 3.4078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19512166157364846,
+      "step": 11890
+    },
+    {
+      "epoch": 0.2975,
+      "grad_norm": 29.0,
+      "grad_norm_var": 4.71875,
+      "learning_rate": 0.0001,
+      "loss": 7.3659,
+      "loss/crossentropy": 2.068368895351887,
+      "loss/hidden": 3.427734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22229769406840205,
+      "step": 11900
+    },
+    {
+      "epoch": 0.29775,
+      "grad_norm": 28.25,
+      "grad_norm_var": 30.575,
+      "learning_rate": 0.0001,
+      "loss": 7.3459,
+      "loss/crossentropy": 1.9633590430021286,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1919392876327038,
+      "step": 11910
+    },
+    {
+      "epoch": 0.298,
+      "grad_norm": 31.5,
+      "grad_norm_var": 25.169205729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2575,
+      "loss/crossentropy": 2.0932082675397394,
+      "loss/hidden": 3.3484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18578519038856028,
+      "step": 11920
+    },
+    {
+      "epoch": 0.29825,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.9197916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3554,
+      "loss/crossentropy": 2.0443739868700503,
+      "loss/hidden": 3.39140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18646586760878564,
+      "step": 11930
+    },
+    {
+      "epoch": 0.2985,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.1629557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3341,
+      "loss/crossentropy": 2.0645432278513907,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20268544293940066,
+      "step": 11940
+    },
+    {
+      "epoch": 0.29875,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.085416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2596,
+      "loss/crossentropy": 1.8821386724710465,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17970443107187747,
+      "step": 11950
+    },
+    {
+      "epoch": 0.299,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.4072265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3139,
+      "loss/crossentropy": 2.1848059430718423,
+      "loss/hidden": 3.311328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19258468672633172,
+      "step": 11960
+    },
+    {
+      "epoch": 0.29925,
+      "grad_norm": 28.625,
+      "grad_norm_var": 1.49140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3064,
+      "loss/crossentropy": 1.9953081101179122,
+      "loss/hidden": 3.341015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17766128927469255,
+      "step": 11970
+    },
+    {
+      "epoch": 0.2995,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.527083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3213,
+      "loss/crossentropy": 1.9512401312589644,
+      "loss/hidden": 3.257421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16765224877744914,
+      "step": 11980
+    },
+    {
+      "epoch": 0.29975,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.8322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3373,
+      "loss/crossentropy": 2.134880567342043,
+      "loss/hidden": 3.353515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19845206197351217,
+      "step": 11990
+    },
+    {
+      "epoch": 0.3,
+      "grad_norm": 29.0,
+      "grad_norm_var": 1.88515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3355,
+      "loss/crossentropy": 2.084690621495247,
+      "loss/hidden": 3.325,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18881614562124013,
+      "step": 12000
+    },
+    {
+      "epoch": 0.30025,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.678125,
+      "learning_rate": 0.0001,
+      "loss": 7.2797,
+      "loss/crossentropy": 2.0142040476202965,
+      "loss/hidden": 3.277734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17630351725965737,
+      "step": 12010
+    },
+    {
+      "epoch": 0.3005,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.0702473958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3524,
+      "loss/crossentropy": 2.0099969252943994,
+      "loss/hidden": 3.3703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17516886200755835,
+      "step": 12020
+    },
+    {
+      "epoch": 0.30075,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.1509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4081,
+      "loss/crossentropy": 2.0550965458154677,
+      "loss/hidden": 3.33359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18450612868182362,
+      "step": 12030
+    },
+    {
+      "epoch": 0.301,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.088997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3732,
+      "loss/crossentropy": 2.015381334722042,
+      "loss/hidden": 3.434375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19866329655051232,
+      "step": 12040
+    },
+    {
+      "epoch": 0.30125,
+      "grad_norm": 32.5,
+      "grad_norm_var": 28.125455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3689,
+      "loss/crossentropy": 2.0619105845689774,
+      "loss/hidden": 3.41640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19213079251348972,
+      "step": 12050
+    },
+    {
+      "epoch": 0.3015,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.5869140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3533,
+      "loss/crossentropy": 2.059162912517786,
+      "loss/hidden": 3.3625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1831585830077529,
+      "step": 12060
+    },
+    {
+      "epoch": 0.30175,
+      "grad_norm": 30.375,
+      "grad_norm_var": 6.6150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.339,
+      "loss/crossentropy": 2.0344214349985124,
+      "loss/hidden": 3.3328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18243852034211158,
+      "step": 12070
+    },
+    {
+      "epoch": 0.302,
+      "grad_norm": 31.375,
+      "grad_norm_var": 10.05390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3457,
+      "loss/crossentropy": 2.222717672586441,
+      "loss/hidden": 3.28671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19160165078938007,
+      "step": 12080
+    },
+    {
+      "epoch": 0.30225,
+      "grad_norm": 30.375,
+      "grad_norm_var": 8.060416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2928,
+      "loss/crossentropy": 1.9902059100568295,
+      "loss/hidden": 3.3921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1817794761620462,
+      "step": 12090
+    },
+    {
+      "epoch": 0.3025,
+      "grad_norm": 29.0,
+      "grad_norm_var": 4.048958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.403,
+      "loss/crossentropy": 1.9990896947681904,
+      "loss/hidden": 3.330859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17872856128960848,
+      "step": 12100
+    },
+    {
+      "epoch": 0.30275,
+      "grad_norm": 32.0,
+      "grad_norm_var": 14.486393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3774,
+      "loss/crossentropy": 2.178261125087738,
+      "loss/hidden": 3.3796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19976076446473598,
+      "step": 12110
+    },
+    {
+      "epoch": 0.303,
+      "grad_norm": 29.0,
+      "grad_norm_var": 25.053059895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.3852,
+      "loss/crossentropy": 2.1738795921206475,
+      "loss/hidden": 3.21171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1802441133186221,
+      "step": 12120
+    },
+    {
+      "epoch": 0.30325,
+      "grad_norm": 32.5,
+      "grad_norm_var": 27.863541666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3119,
+      "loss/crossentropy": 2.017075891792774,
+      "loss/hidden": 3.39296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1819239752367139,
+      "step": 12130
+    },
+    {
+      "epoch": 0.3035,
+      "grad_norm": 5704253440.0,
+      "grad_norm_var": 2.033656683216328e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3187,
+      "loss/crossentropy": 2.1378798320889474,
+      "loss/hidden": 3.403125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18621215373277664,
+      "step": 12140
+    },
+    {
+      "epoch": 0.30375,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.0336566834599473e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3375,
+      "loss/crossentropy": 2.0977103441953657,
+      "loss/hidden": 3.33046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18937066961079835,
+      "step": 12150
+    },
+    {
+      "epoch": 0.304,
+      "grad_norm": 30.0,
+      "grad_norm_var": 25.7119140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2813,
+      "loss/crossentropy": 2.0422945946455,
+      "loss/hidden": 3.29453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17747708465903997,
+      "step": 12160
+    },
+    {
+      "epoch": 0.30425,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.8973307291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2982,
+      "loss/crossentropy": 2.1438152998685838,
+      "loss/hidden": 3.4,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19225314557552337,
+      "step": 12170
+    },
+    {
+      "epoch": 0.3045,
+      "grad_norm": 32.5,
+      "grad_norm_var": 6.688997395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2886,
+      "loss/crossentropy": 2.033204630762339,
+      "loss/hidden": 3.326953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1741345826536417,
+      "step": 12180
+    },
+    {
+      "epoch": 0.30475,
+      "grad_norm": 28.5,
+      "grad_norm_var": 7.262434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3812,
+      "loss/crossentropy": 1.9149722829461098,
+      "loss/hidden": 3.384765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16993715493008493,
+      "step": 12190
+    },
+    {
+      "epoch": 0.305,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.315625,
+      "learning_rate": 0.0001,
+      "loss": 7.355,
+      "loss/crossentropy": 2.200256870687008,
+      "loss/hidden": 3.3703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1957419654354453,
+      "step": 12200
+    },
+    {
+      "epoch": 0.30525,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.8291666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2689,
+      "loss/crossentropy": 1.9441036701202392,
+      "loss/hidden": 3.450390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19158205185085536,
+      "step": 12210
+    },
+    {
+      "epoch": 0.3055,
+      "grad_norm": 32.5,
+      "grad_norm_var": 224.99993489583332,
+      "learning_rate": 0.0001,
+      "loss": 7.4205,
+      "loss/crossentropy": 2.21477717012167,
+      "loss/hidden": 3.353515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19972570687532426,
+      "step": 12220
+    },
+    {
+      "epoch": 0.30575,
+      "grad_norm": 77.0,
+      "grad_norm_var": 161.4994140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3221,
+      "loss/crossentropy": 2.243621972203255,
+      "loss/hidden": 3.381640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20230026841163634,
+      "step": 12230
+    },
+    {
+      "epoch": 0.306,
+      "grad_norm": 29.125,
+      "grad_norm_var": 136.2509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2544,
+      "loss/crossentropy": 2.1267666652798654,
+      "loss/hidden": 3.30078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18361538834869862,
+      "step": 12240
+    },
+    {
+      "epoch": 0.30625,
+      "grad_norm": 43.0,
+      "grad_norm_var": 15.1384765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3457,
+      "loss/crossentropy": 2.097319718450308,
+      "loss/hidden": 3.380859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1941752176731825,
+      "step": 12250
+    },
+    {
+      "epoch": 0.3065,
+      "grad_norm": 27.125,
+      "grad_norm_var": 45.91875,
+      "learning_rate": 0.0001,
+      "loss": 7.1913,
+      "loss/crossentropy": 2.024158325791359,
+      "loss/hidden": 3.301171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18036661259829997,
+      "step": 12260
+    },
+    {
+      "epoch": 0.30675,
+      "grad_norm": 32.5,
+      "grad_norm_var": 34.94765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2604,
+      "loss/crossentropy": 2.05818811878562,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19272202141582967,
+      "step": 12270
+    },
+    {
+      "epoch": 0.307,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.7958333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3478,
+      "loss/crossentropy": 2.058871729671955,
+      "loss/hidden": 3.316015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1805756026878953,
+      "step": 12280
+    },
+    {
+      "epoch": 0.30725,
+      "grad_norm": 37.0,
+      "grad_norm_var": 4.418489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3202,
+      "loss/crossentropy": 1.9400940239429474,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19434008933603764,
+      "step": 12290
+    },
+    {
+      "epoch": 0.3075,
+      "grad_norm": 30.5,
+      "grad_norm_var": 30.1572265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2939,
+      "loss/crossentropy": 2.069047340750694,
+      "loss/hidden": 3.42421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20558738969266416,
+      "step": 12300
+    },
+    {
+      "epoch": 0.30775,
+      "grad_norm": 33.0,
+      "grad_norm_var": 11.9666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4109,
+      "loss/crossentropy": 2.227694994211197,
+      "loss/hidden": 3.395703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18980047646909953,
+      "step": 12310
+    },
+    {
+      "epoch": 0.308,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.90390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3686,
+      "loss/crossentropy": 2.0309864141047003,
+      "loss/hidden": 3.42578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1816682495176792,
+      "step": 12320
+    },
+    {
+      "epoch": 0.30825,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.7249348958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4105,
+      "loss/crossentropy": 2.2505668699741364,
+      "loss/hidden": 3.307421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19684594608843325,
+      "step": 12330
+    },
+    {
+      "epoch": 0.3085,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.9535807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3679,
+      "loss/crossentropy": 2.0422501020133494,
+      "loss/hidden": 3.223828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1723391016945243,
+      "step": 12340
+    },
+    {
+      "epoch": 0.30875,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.9186848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3186,
+      "loss/crossentropy": 2.190096604824066,
+      "loss/hidden": 3.299609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1794669708237052,
+      "step": 12350
+    },
+    {
+      "epoch": 0.309,
+      "grad_norm": 30.0,
+      "grad_norm_var": 1.7926432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2068,
+      "loss/crossentropy": 2.011892383173108,
+      "loss/hidden": 3.3046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17626657225191594,
+      "step": 12360
+    },
+    {
+      "epoch": 0.30925,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.3900390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3963,
+      "loss/crossentropy": 2.1367972552776338,
+      "loss/hidden": 3.356640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19975323602557182,
+      "step": 12370
+    },
+    {
+      "epoch": 0.3095,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.9785807291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3586,
+      "loss/crossentropy": 2.1065737903118134,
+      "loss/hidden": 3.3296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1861647253856063,
+      "step": 12380
+    },
+    {
+      "epoch": 0.30975,
+      "grad_norm": 34.0,
+      "grad_norm_var": 49.94583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4425,
+      "loss/crossentropy": 2.047655486315489,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19645441174507142,
+      "step": 12390
+    },
+    {
+      "epoch": 0.31,
+      "grad_norm": 30.75,
+      "grad_norm_var": 55.521809895833336,
+      "learning_rate": 0.0001,
+      "loss": 7.3384,
+      "loss/crossentropy": 2.1421022772789002,
+      "loss/hidden": 3.453515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21594423688948156,
+      "step": 12400
+    },
+    {
+      "epoch": 0.31025,
+      "grad_norm": 28.625,
+      "grad_norm_var": 3.630989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4915,
+      "loss/crossentropy": 1.9852606005966664,
+      "loss/hidden": 3.3875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1970728723332286,
+      "step": 12410
+    },
+    {
+      "epoch": 0.3105,
+      "grad_norm": 30.375,
+      "grad_norm_var": 13.821875,
+      "learning_rate": 0.0001,
+      "loss": 7.316,
+      "loss/crossentropy": 2.0221208080649378,
+      "loss/hidden": 3.40546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18263351675122977,
+      "step": 12420
+    },
+    {
+      "epoch": 0.31075,
+      "grad_norm": 27.75,
+      "grad_norm_var": 3.7905598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4569,
+      "loss/crossentropy": 2.2214868292212486,
+      "loss/hidden": 3.430078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20803299229592084,
+      "step": 12430
+    },
+    {
+      "epoch": 0.311,
+      "grad_norm": 28.0,
+      "grad_norm_var": 5.039322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2246,
+      "loss/crossentropy": 1.968916893005371,
+      "loss/hidden": 3.398046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1823066620156169,
+      "step": 12440
+    },
+    {
+      "epoch": 0.31125,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.011458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2986,
+      "loss/crossentropy": 2.102560856938362,
+      "loss/hidden": 3.462109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19421595316380263,
+      "step": 12450
+    },
+    {
+      "epoch": 0.3115,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.7122395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4102,
+      "loss/crossentropy": 1.8904587842524052,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16765992902219296,
+      "step": 12460
+    },
+    {
+      "epoch": 0.31175,
+      "grad_norm": 32.75,
+      "grad_norm_var": 1.3614583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.317,
+      "loss/crossentropy": 2.2456528916954994,
+      "loss/hidden": 3.3265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18445245549082756,
+      "step": 12470
+    },
+    {
+      "epoch": 0.312,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.9613932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2527,
+      "loss/crossentropy": 1.9254849448800087,
+      "loss/hidden": 3.330859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17210696106776596,
+      "step": 12480
+    },
+    {
+      "epoch": 0.31225,
+      "grad_norm": 29.0,
+      "grad_norm_var": 2.503059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3589,
+      "loss/crossentropy": 1.963121373206377,
+      "loss/hidden": 3.3421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17580699175596237,
+      "step": 12490
+    },
+    {
+      "epoch": 0.3125,
+      "grad_norm": 27.375,
+      "grad_norm_var": 2.6510416666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3302,
+      "loss/crossentropy": 2.0668979212641716,
+      "loss/hidden": 3.33828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1756731817498803,
+      "step": 12500
+    },
+    {
+      "epoch": 0.31275,
+      "grad_norm": 32.5,
+      "grad_norm_var": 9.658072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.396,
+      "loss/crossentropy": 2.0963846892118454,
+      "loss/hidden": 3.37578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19187606330960988,
+      "step": 12510
+    },
+    {
+      "epoch": 0.313,
+      "grad_norm": 28.125,
+      "grad_norm_var": 8.387239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.291,
+      "loss/crossentropy": 2.1238081738352776,
+      "loss/hidden": 3.319921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18899311302229763,
+      "step": 12520
+    },
+    {
+      "epoch": 0.31325,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.2712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2927,
+      "loss/crossentropy": 1.8629853092133999,
+      "loss/hidden": 3.430078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18793845577165486,
+      "step": 12530
+    },
+    {
+      "epoch": 0.3135,
+      "grad_norm": 31.25,
+      "grad_norm_var": 29.926041666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.334,
+      "loss/crossentropy": 2.1922558069229128,
+      "loss/hidden": 3.371484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18994415253400804,
+      "step": 12540
+    },
+    {
+      "epoch": 0.31375,
+      "grad_norm": 33.0,
+      "grad_norm_var": 19.626822916666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4246,
+      "loss/crossentropy": 2.0747738771140574,
+      "loss/hidden": 3.463671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1956952316686511,
+      "step": 12550
+    },
+    {
+      "epoch": 0.314,
+      "grad_norm": 31.0,
+      "grad_norm_var": 5.9822265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3921,
+      "loss/crossentropy": 2.168559101223946,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19871533028781413,
+      "step": 12560
+    },
+    {
+      "epoch": 0.31425,
+      "grad_norm": 29.875,
+      "grad_norm_var": 23.478125,
+      "learning_rate": 0.0001,
+      "loss": 7.381,
+      "loss/crossentropy": 2.1687012270092962,
+      "loss/hidden": 3.2828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17819978184998037,
+      "step": 12570
+    },
+    {
+      "epoch": 0.3145,
+      "grad_norm": 30.5,
+      "grad_norm_var": 24.369791666666668,
+      "learning_rate": 0.0001,
+      "loss": 7.2273,
+      "loss/crossentropy": 2.0223761543631555,
+      "loss/hidden": 3.2953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17995049208402633,
+      "step": 12580
+    },
+    {
+      "epoch": 0.31475,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.8067057291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3672,
+      "loss/crossentropy": 2.056264813989401,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19224987253546716,
+      "step": 12590
+    },
+    {
+      "epoch": 0.315,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.3018229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2404,
+      "loss/crossentropy": 2.0127123326063154,
+      "loss/hidden": 3.36640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17083216030150652,
+      "step": 12600
+    },
+    {
+      "epoch": 0.31525,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.94375,
+      "learning_rate": 0.0001,
+      "loss": 7.3918,
+      "loss/crossentropy": 2.0961378000676634,
+      "loss/hidden": 3.3796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18820952698588372,
+      "step": 12610
+    },
+    {
+      "epoch": 0.3155,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.910416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2715,
+      "loss/crossentropy": 2.1737264052033423,
+      "loss/hidden": 3.33359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1752359176054597,
+      "step": 12620
+    },
+    {
+      "epoch": 0.31575,
+      "grad_norm": 28.5,
+      "grad_norm_var": 12.914322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2866,
+      "loss/crossentropy": 2.04836600497365,
+      "loss/hidden": 3.308984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1814719710499048,
+      "step": 12630
+    },
+    {
+      "epoch": 0.316,
+      "grad_norm": 32.5,
+      "grad_norm_var": 5.895572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4359,
+      "loss/crossentropy": 2.0633833378553392,
+      "loss/hidden": 3.38515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1844972724094987,
+      "step": 12640
+    },
+    {
+      "epoch": 0.31625,
+      "grad_norm": 29.5,
+      "grad_norm_var": 4.556705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3404,
+      "loss/crossentropy": 2.1031613536179066,
+      "loss/hidden": 3.38671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1890261113177985,
+      "step": 12650
+    },
+    {
+      "epoch": 0.3165,
+      "grad_norm": 33.25,
+      "grad_norm_var": 7.43515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3116,
+      "loss/crossentropy": 2.0913070663809776,
+      "loss/hidden": 3.48515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19625448603183032,
+      "step": 12660
+    },
+    {
+      "epoch": 0.31675,
+      "grad_norm": 30.0,
+      "grad_norm_var": 24.72265625,
+      "learning_rate": 0.0001,
+      "loss": 7.309,
+      "loss/crossentropy": 2.1766112834215163,
+      "loss/hidden": 3.398046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19453570377081633,
+      "step": 12670
+    },
+    {
+      "epoch": 0.317,
+      "grad_norm": 28.0,
+      "grad_norm_var": 3.5518229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2193,
+      "loss/crossentropy": 2.1292213678359984,
+      "loss/hidden": 3.312890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17923556556925177,
+      "step": 12680
+    },
+    {
+      "epoch": 0.31725,
+      "grad_norm": 33.75,
+      "grad_norm_var": 3.2041666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3568,
+      "loss/crossentropy": 2.0966979548335076,
+      "loss/hidden": 3.34375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1796438904479146,
+      "step": 12690
+    },
+    {
+      "epoch": 0.3175,
+      "grad_norm": 30.125,
+      "grad_norm_var": 21.9853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3281,
+      "loss/crossentropy": 2.2050976656377315,
+      "loss/hidden": 3.317578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18696724623441696,
+      "step": 12700
+    },
+    {
+      "epoch": 0.31775,
+      "grad_norm": 31.25,
+      "grad_norm_var": 20.936458333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2583,
+      "loss/crossentropy": 2.0703148849308493,
+      "loss/hidden": 3.267578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17689838781952857,
+      "step": 12710
+    },
+    {
+      "epoch": 0.318,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.9395833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3007,
+      "loss/crossentropy": 1.9791635520756246,
+      "loss/hidden": 3.359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17443582694977522,
+      "step": 12720
+    },
+    {
+      "epoch": 0.31825,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.5155598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2557,
+      "loss/crossentropy": 2.0753975957632065,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18456004988402128,
+      "step": 12730
+    },
+    {
+      "epoch": 0.3185,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.3518229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2658,
+      "loss/crossentropy": 2.0314756602048876,
+      "loss/hidden": 3.3890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18754468094557525,
+      "step": 12740
+    },
+    {
+      "epoch": 0.31875,
+      "grad_norm": 29.25,
+      "grad_norm_var": 1.6634765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2545,
+      "loss/crossentropy": 2.027373602986336,
+      "loss/hidden": 3.366796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18138464018702508,
+      "step": 12750
+    },
+    {
+      "epoch": 0.319,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.2837890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2711,
+      "loss/crossentropy": 2.0996690608561037,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19384829625487326,
+      "step": 12760
+    },
+    {
+      "epoch": 0.31925,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.7580729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3448,
+      "loss/crossentropy": 2.081255576014519,
+      "loss/hidden": 3.48515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2054756512865424,
+      "step": 12770
+    },
+    {
+      "epoch": 0.3195,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.6337890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3614,
+      "loss/crossentropy": 2.109497997164726,
+      "loss/hidden": 3.2984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1826931856572628,
+      "step": 12780
+    },
+    {
+      "epoch": 0.31975,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.9125,
+      "learning_rate": 0.0001,
+      "loss": 7.3569,
+      "loss/crossentropy": 2.071355660259724,
+      "loss/hidden": 3.390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18764556515961886,
+      "step": 12790
+    },
+    {
+      "epoch": 0.32,
+      "grad_norm": 31.0,
+      "grad_norm_var": 5.68125,
+      "learning_rate": 0.0001,
+      "loss": 7.3399,
+      "loss/crossentropy": 1.9978457435965538,
+      "loss/hidden": 3.365625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17209083335474135,
+      "step": 12800
+    },
+    {
+      "epoch": 0.32025,
+      "grad_norm": 28.75,
+      "grad_norm_var": 20.754166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3016,
+      "loss/crossentropy": 1.984939170628786,
+      "loss/hidden": 3.50625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19444225933402776,
+      "step": 12810
+    },
+    {
+      "epoch": 0.3205,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.3942057291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3066,
+      "loss/crossentropy": 1.8981015786528588,
+      "loss/hidden": 3.41171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18696952695026994,
+      "step": 12820
+    },
+    {
+      "epoch": 0.32075,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.9218098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3974,
+      "loss/crossentropy": 2.1415200501680376,
+      "loss/hidden": 3.434375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20154954344034196,
+      "step": 12830
+    },
+    {
+      "epoch": 0.321,
+      "grad_norm": 31.375,
+      "grad_norm_var": 35.8400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3048,
+      "loss/crossentropy": 2.1483262166380883,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19059547781944275,
+      "step": 12840
+    },
+    {
+      "epoch": 0.32125,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.561393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4233,
+      "loss/crossentropy": 2.1605855494737627,
+      "loss/hidden": 3.3015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19197470052167773,
+      "step": 12850
+    },
+    {
+      "epoch": 0.3215,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.98125,
+      "learning_rate": 0.0001,
+      "loss": 7.3826,
+      "loss/crossentropy": 2.1228017687797545,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21111836601048709,
+      "step": 12860
+    },
+    {
+      "epoch": 0.32175,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.6666666666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.317,
+      "loss/crossentropy": 1.9638956546783448,
+      "loss/hidden": 3.3625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18815859649330377,
+      "step": 12870
+    },
+    {
+      "epoch": 0.322,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.5927083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3412,
+      "loss/crossentropy": 2.15369998216629,
+      "loss/hidden": 3.32109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19398804232478142,
+      "step": 12880
+    },
+    {
+      "epoch": 0.32225,
+      "grad_norm": 28.875,
+      "grad_norm_var": 4.522916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3785,
+      "loss/crossentropy": 2.134490595757961,
+      "loss/hidden": 3.367578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19743212331086396,
+      "step": 12890
+    },
+    {
+      "epoch": 0.3225,
+      "grad_norm": 28.125,
+      "grad_norm_var": 4.824739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2795,
+      "loss/crossentropy": 2.05883831679821,
+      "loss/hidden": 3.38515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1887126075103879,
+      "step": 12900
+    },
+    {
+      "epoch": 0.32275,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.6257994456428096e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3637,
+      "loss/crossentropy": 1.865815930068493,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16476531345397233,
+      "step": 12910
+    },
+    {
+      "epoch": 0.323,
+      "grad_norm": 27.375,
+      "grad_norm_var": 1.6257994456056202e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3231,
+      "loss/crossentropy": 2.1335863292217256,
+      "loss/hidden": 3.3234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18054731655865908,
+      "step": 12920
+    },
+    {
+      "epoch": 0.32325,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.6822265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2992,
+      "loss/crossentropy": 2.029070366919041,
+      "loss/hidden": 3.292578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1656917490065098,
+      "step": 12930
+    },
+    {
+      "epoch": 0.3235,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.9139973958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3529,
+      "loss/crossentropy": 2.0367226876318454,
+      "loss/hidden": 3.446484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18601053059101105,
+      "step": 12940
+    },
+    {
+      "epoch": 0.32375,
+      "grad_norm": 27.875,
+      "grad_norm_var": 2.1228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.146,
+      "loss/crossentropy": 2.008280509710312,
+      "loss/hidden": 3.339453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17591428142040968,
+      "step": 12950
+    },
+    {
+      "epoch": 0.324,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.8893229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3616,
+      "loss/crossentropy": 2.1186923176050185,
+      "loss/hidden": 3.476171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20130661278963088,
+      "step": 12960
+    },
+    {
+      "epoch": 0.32425,
+      "grad_norm": 28.0,
+      "grad_norm_var": 1.4749348958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3933,
+      "loss/crossentropy": 2.1367026805877685,
+      "loss/hidden": 3.41640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19593327604234217,
+      "step": 12970
+    },
+    {
+      "epoch": 0.3245,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.5009765625,
+      "learning_rate": 0.0001,
+      "loss": 7.36,
+      "loss/crossentropy": 1.9709487736225129,
+      "loss/hidden": 3.388671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18522804593667389,
+      "step": 12980
+    },
+    {
+      "epoch": 0.32475,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.037955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3995,
+      "loss/crossentropy": 2.182401825487614,
+      "loss/hidden": 3.288671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1944254267960787,
+      "step": 12990
+    },
+    {
+      "epoch": 0.325,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.8259765625,
+      "learning_rate": 0.0001,
+      "loss": 7.297,
+      "loss/crossentropy": 2.144014260172844,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18785431496798993,
+      "step": 13000
+    },
+    {
+      "epoch": 0.32525,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.5858723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4172,
+      "loss/crossentropy": 2.185704696178436,
+      "loss/hidden": 3.476953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21230035796761512,
+      "step": 13010
+    },
+    {
+      "epoch": 0.3255,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.0989583333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3542,
+      "loss/crossentropy": 2.169140038639307,
+      "loss/hidden": 3.367578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20137128187343478,
+      "step": 13020
+    },
+    {
+      "epoch": 0.32575,
+      "grad_norm": 29.0,
+      "grad_norm_var": 1.7613932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3018,
+      "loss/crossentropy": 2.0429045438766478,
+      "loss/hidden": 3.306640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1801679054275155,
+      "step": 13030
+    },
+    {
+      "epoch": 0.326,
+      "grad_norm": 28.25,
+      "grad_norm_var": 20.3603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3076,
+      "loss/crossentropy": 2.090936814248562,
+      "loss/hidden": 3.363671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18894913028925658,
+      "step": 13040
+    },
+    {
+      "epoch": 0.32625,
+      "grad_norm": 29.5,
+      "grad_norm_var": 21.5759765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3032,
+      "loss/crossentropy": 2.0933708012104035,
+      "loss/hidden": 3.306640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18349476121366023,
+      "step": 13050
+    },
+    {
+      "epoch": 0.3265,
+      "grad_norm": 31.125,
+      "grad_norm_var": 23.737434895833335,
+      "learning_rate": 0.0001,
+      "loss": 7.4007,
+      "loss/crossentropy": 1.9636264845728875,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17653203681111335,
+      "step": 13060
+    },
+    {
+      "epoch": 0.32675,
+      "grad_norm": 30.875,
+      "grad_norm_var": 26.170833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2685,
+      "loss/crossentropy": 2.0459864370524885,
+      "loss/hidden": 3.392578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18335657650604845,
+      "step": 13070
+    },
+    {
+      "epoch": 0.327,
+      "grad_norm": 32.25,
+      "grad_norm_var": 13.033072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3321,
+      "loss/crossentropy": 2.2077785804867744,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1968739289790392,
+      "step": 13080
+    },
+    {
+      "epoch": 0.32725,
+      "grad_norm": 29.625,
+      "grad_norm_var": 8.8634765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2613,
+      "loss/crossentropy": 2.039609357714653,
+      "loss/hidden": 3.45078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19310947302728892,
+      "step": 13090
+    },
+    {
+      "epoch": 0.3275,
+      "grad_norm": 29.875,
+      "grad_norm_var": 8.334309895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3587,
+      "loss/crossentropy": 2.0822102136909963,
+      "loss/hidden": 3.26640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1863661216571927,
+      "step": 13100
+    },
+    {
+      "epoch": 0.32775,
+      "grad_norm": 34.25,
+      "grad_norm_var": 8.694205729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3602,
+      "loss/crossentropy": 2.167416235804558,
+      "loss/hidden": 3.36484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19521966725587844,
+      "step": 13110
+    },
+    {
+      "epoch": 0.328,
+      "grad_norm": 29.125,
+      "grad_norm_var": 7.191080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3767,
+      "loss/crossentropy": 2.07031906619668,
+      "loss/hidden": 3.4015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22052818778902292,
+      "step": 13120
+    },
+    {
+      "epoch": 0.32825,
+      "grad_norm": 35.5,
+      "grad_norm_var": 5.821809895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5034,
+      "loss/crossentropy": 2.1318026900291445,
+      "loss/hidden": 3.4046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19695397429168224,
+      "step": 13130
+    },
+    {
+      "epoch": 0.3285,
+      "grad_norm": 29.5,
+      "grad_norm_var": 4.738541666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2907,
+      "loss/crossentropy": 1.893085064738989,
+      "loss/hidden": 3.421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17454516123980285,
+      "step": 13140
+    },
+    {
+      "epoch": 0.32875,
+      "grad_norm": 33.25,
+      "grad_norm_var": 4.96640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3911,
+      "loss/crossentropy": 2.0709325544536115,
+      "loss/hidden": 3.405078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19108830243349076,
+      "step": 13150
+    },
+    {
+      "epoch": 0.329,
+      "grad_norm": 30.375,
+      "grad_norm_var": 6.593489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.278,
+      "loss/crossentropy": 2.0809618443250657,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1867564545944333,
+      "step": 13160
+    },
+    {
+      "epoch": 0.32925,
+      "grad_norm": 28.875,
+      "grad_norm_var": 5.355143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.1865,
+      "loss/crossentropy": 1.86175979077816,
+      "loss/hidden": 3.41640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18658734802156687,
+      "step": 13170
+    },
+    {
+      "epoch": 0.3295,
+      "grad_norm": 32.5,
+      "grad_norm_var": 7.297916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2993,
+      "loss/crossentropy": 2.0688765406608582,
+      "loss/hidden": 3.320703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1887798959389329,
+      "step": 13180
+    },
+    {
+      "epoch": 0.32975,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.1462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.396,
+      "loss/crossentropy": 2.1349074259400367,
+      "loss/hidden": 3.3734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1891916124150157,
+      "step": 13190
+    },
+    {
+      "epoch": 0.33,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.2080729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2194,
+      "loss/crossentropy": 2.0701663225889204,
+      "loss/hidden": 3.290625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19238610491156577,
+      "step": 13200
+    },
+    {
+      "epoch": 0.33025,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.3447265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3476,
+      "loss/crossentropy": 2.126928760111332,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18552705701440572,
+      "step": 13210
+    },
+    {
+      "epoch": 0.3305,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.4916666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3275,
+      "loss/crossentropy": 2.0960667990148067,
+      "loss/hidden": 3.3546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18840519580990076,
+      "step": 13220
+    },
+    {
+      "epoch": 0.33075,
+      "grad_norm": 29.0,
+      "grad_norm_var": 31.097916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3449,
+      "loss/crossentropy": 1.9724566139280797,
+      "loss/hidden": 3.401171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17257819082587958,
+      "step": 13230
+    },
+    {
+      "epoch": 0.331,
+      "grad_norm": 30.25,
+      "grad_norm_var": 29.516080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5214,
+      "loss/crossentropy": 2.144184100627899,
+      "loss/hidden": 3.4296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19250041600316764,
+      "step": 13240
+    },
+    {
+      "epoch": 0.33125,
+      "grad_norm": 28.25,
+      "grad_norm_var": 6.55,
+      "learning_rate": 0.0001,
+      "loss": 7.3245,
+      "loss/crossentropy": 1.9724821582436562,
+      "loss/hidden": 3.423046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18454758413136005,
+      "step": 13250
+    },
+    {
+      "epoch": 0.3315,
+      "grad_norm": 35.75,
+      "grad_norm_var": 10.574739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2067,
+      "loss/crossentropy": 2.0341859996318816,
+      "loss/hidden": 3.466015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19355686828494073,
+      "step": 13260
+    },
+    {
+      "epoch": 0.33175,
+      "grad_norm": 31.875,
+      "grad_norm_var": 6.093489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3386,
+      "loss/crossentropy": 2.1593449860811234,
+      "loss/hidden": 3.282421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18228702070191502,
+      "step": 13270
+    },
+    {
+      "epoch": 0.332,
+      "grad_norm": 32.25,
+      "grad_norm_var": 4.518489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2584,
+      "loss/crossentropy": 2.021310421824455,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.193563433829695,
+      "step": 13280
+    },
+    {
+      "epoch": 0.33225,
+      "grad_norm": 33.25,
+      "grad_norm_var": 3.008072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3654,
+      "loss/crossentropy": 2.010246267169714,
+      "loss/hidden": 3.398046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18375244587659836,
+      "step": 13290
+    },
+    {
+      "epoch": 0.3325,
+      "grad_norm": 33.5,
+      "grad_norm_var": 1.8309895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4013,
+      "loss/crossentropy": 2.1324703454971314,
+      "loss/hidden": 3.42109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19651575423777104,
+      "step": 13300
+    },
+    {
+      "epoch": 0.33275,
+      "grad_norm": 31.625,
+      "grad_norm_var": 9.018489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4123,
+      "loss/crossentropy": 2.2024870067834854,
+      "loss/hidden": 3.36484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19117040373384953,
+      "step": 13310
+    },
+    {
+      "epoch": 0.333,
+      "grad_norm": 29.5,
+      "grad_norm_var": 54.362239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3561,
+      "loss/crossentropy": 1.9291939452290534,
+      "loss/hidden": 3.33828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18055410776287317,
+      "step": 13320
+    },
+    {
+      "epoch": 0.33325,
+      "grad_norm": 30.5,
+      "grad_norm_var": 53.271875,
+      "learning_rate": 0.0001,
+      "loss": 7.3417,
+      "loss/crossentropy": 1.9812588825821877,
+      "loss/hidden": 3.246875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17277830513194203,
+      "step": 13330
+    },
+    {
+      "epoch": 0.3335,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.0332682291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4792,
+      "loss/crossentropy": 2.1468162171542646,
+      "loss/hidden": 3.29453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1815652133896947,
+      "step": 13340
+    },
+    {
+      "epoch": 0.33375,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.8942057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3437,
+      "loss/crossentropy": 1.9899487346410751,
+      "loss/hidden": 3.336328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1736677044071257,
+      "step": 13350
+    },
+    {
+      "epoch": 0.334,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.374955311060432e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4592,
+      "loss/crossentropy": 2.0041578873991965,
+      "loss/hidden": 3.38984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19686540886759757,
+      "step": 13360
+    },
+    {
+      "epoch": 0.33425,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.3749553110757407e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3521,
+      "loss/crossentropy": 2.050678627192974,
+      "loss/hidden": 3.35859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1835683614946902,
+      "step": 13370
+    },
+    {
+      "epoch": 0.3345,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.4561848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3463,
+      "loss/crossentropy": 2.255459001660347,
+      "loss/hidden": 3.3484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2026721488684416,
+      "step": 13380
+    },
+    {
+      "epoch": 0.33475,
+      "grad_norm": 55.0,
+      "grad_norm_var": 37.805989583333336,
+      "learning_rate": 0.0001,
+      "loss": 7.432,
+      "loss/crossentropy": 2.092339722812176,
+      "loss/hidden": 3.445703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20365451760590075,
+      "step": 13390
+    },
+    {
+      "epoch": 0.335,
+      "grad_norm": 29.5,
+      "grad_norm_var": 39.1962890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3082,
+      "loss/crossentropy": 2.060484157502651,
+      "loss/hidden": 3.49921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19208470694720745,
+      "step": 13400
+    },
+    {
+      "epoch": 0.33525,
+      "grad_norm": 30.875,
+      "grad_norm_var": 4.119205729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3159,
+      "loss/crossentropy": 2.0885482341051103,
+      "loss/hidden": 3.3484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18701701434329152,
+      "step": 13410
+    },
+    {
+      "epoch": 0.3355,
+      "grad_norm": 34.75,
+      "grad_norm_var": 3.8354166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3792,
+      "loss/crossentropy": 2.064309825748205,
+      "loss/hidden": 3.40546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20586481597274542,
+      "step": 13420
+    },
+    {
+      "epoch": 0.33575,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.535416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4193,
+      "loss/crossentropy": 1.8737250491976738,
+      "loss/hidden": 3.427734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19758351668715476,
+      "step": 13430
+    },
+    {
+      "epoch": 0.336,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.8686848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4324,
+      "loss/crossentropy": 1.9396042831242084,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18089414723217487,
+      "step": 13440
+    },
+    {
+      "epoch": 0.33625,
+      "grad_norm": 28.875,
+      "grad_norm_var": 4.540625,
+      "learning_rate": 0.0001,
+      "loss": 7.3798,
+      "loss/crossentropy": 2.004528859257698,
+      "loss/hidden": 3.442578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22120596412569285,
+      "step": 13450
+    },
+    {
+      "epoch": 0.3365,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.69375,
+      "learning_rate": 0.0001,
+      "loss": 7.3708,
+      "loss/crossentropy": 2.170805335044861,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18870076667517424,
+      "step": 13460
+    },
+    {
+      "epoch": 0.33675,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.9284656301175997e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3594,
+      "loss/crossentropy": 2.0726011991500854,
+      "loss/hidden": 3.310546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18316982481628657,
+      "step": 13470
+    },
+    {
+      "epoch": 0.337,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.999739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3072,
+      "loss/crossentropy": 2.0701726540923118,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18232779456302523,
+      "step": 13480
+    },
+    {
+      "epoch": 0.33725,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.01640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3553,
+      "loss/crossentropy": 2.066875821352005,
+      "loss/hidden": 3.34765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18299774192273616,
+      "step": 13490
+    },
+    {
+      "epoch": 0.3375,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.4489583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2708,
+      "loss/crossentropy": 2.110422171652317,
+      "loss/hidden": 3.353125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19538786429911853,
+      "step": 13500
+    },
+    {
+      "epoch": 0.33775,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.06015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3529,
+      "loss/crossentropy": 2.1061682522296907,
+      "loss/hidden": 3.4078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19031856823712587,
+      "step": 13510
+    },
+    {
+      "epoch": 0.338,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.40390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2473,
+      "loss/crossentropy": 2.0029176853597166,
+      "loss/hidden": 3.284375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.182035060133785,
+      "step": 13520
+    },
+    {
+      "epoch": 0.33825,
+      "grad_norm": 29.375,
+      "grad_norm_var": 9.417708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3084,
+      "loss/crossentropy": 2.1238187912851574,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.188486148416996,
+      "step": 13530
+    },
+    {
+      "epoch": 0.3385,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.3983723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2256,
+      "loss/crossentropy": 1.9766217768192291,
+      "loss/hidden": 3.326953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17294846558943391,
+      "step": 13540
+    },
+    {
+      "epoch": 0.33875,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.0874348958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.327,
+      "loss/crossentropy": 2.1272383123636245,
+      "loss/hidden": 3.345703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18510236088186502,
+      "step": 13550
+    },
+    {
+      "epoch": 0.339,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.10390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2885,
+      "loss/crossentropy": 2.0515687823295594,
+      "loss/hidden": 3.368359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18967761769890784,
+      "step": 13560
+    },
+    {
+      "epoch": 0.33925,
+      "grad_norm": 29.625,
+      "grad_norm_var": 11.933072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3573,
+      "loss/crossentropy": 2.0446967758238315,
+      "loss/hidden": 3.425,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18640230242162942,
+      "step": 13570
+    },
+    {
+      "epoch": 0.3395,
+      "grad_norm": 28.875,
+      "grad_norm_var": 1.84765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2965,
+      "loss/crossentropy": 1.9837520524859429,
+      "loss/hidden": 3.3171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1916389312595129,
+      "step": 13580
+    },
+    {
+      "epoch": 0.33975,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.1791666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3632,
+      "loss/crossentropy": 2.1806528866291046,
+      "loss/hidden": 3.337109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19690024815499782,
+      "step": 13590
+    },
+    {
+      "epoch": 0.34,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.3385416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3761,
+      "loss/crossentropy": 2.215220719575882,
+      "loss/hidden": 3.312109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19034639187157154,
+      "step": 13600
+    },
+    {
+      "epoch": 0.34025,
+      "grad_norm": 27.375,
+      "grad_norm_var": 2.7129557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3129,
+      "loss/crossentropy": 2.16262392103672,
+      "loss/hidden": 3.503125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20174582321196793,
+      "step": 13610
+    },
+    {
+      "epoch": 0.3405,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.2363932291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3501,
+      "loss/crossentropy": 2.126379433274269,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18432336021214724,
+      "step": 13620
+    },
+    {
+      "epoch": 0.34075,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.8176432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3601,
+      "loss/crossentropy": 2.0324837125837805,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18331200983375312,
+      "step": 13630
+    },
+    {
+      "epoch": 0.341,
+      "grad_norm": 29.0,
+      "grad_norm_var": 1.7942057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3882,
+      "loss/crossentropy": 2.176556921005249,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19637300558388232,
+      "step": 13640
+    },
+    {
+      "epoch": 0.34125,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.758072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1992,
+      "loss/crossentropy": 1.9680285774171353,
+      "loss/hidden": 3.30546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1837764661759138,
+      "step": 13650
+    },
+    {
+      "epoch": 0.3415,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.6936848958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.361,
+      "loss/crossentropy": 2.1511219844222067,
+      "loss/hidden": 3.29140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1746565790846944,
+      "step": 13660
+    },
+    {
+      "epoch": 0.34175,
+      "grad_norm": 28.375,
+      "grad_norm_var": 2.312239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3856,
+      "loss/crossentropy": 2.187084162980318,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18925763312727212,
+      "step": 13670
+    },
+    {
+      "epoch": 0.342,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.4139973958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.196,
+      "loss/crossentropy": 2.0472760550677775,
+      "loss/hidden": 3.32109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17813790533691645,
+      "step": 13680
+    },
+    {
+      "epoch": 0.34225,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.6458333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4184,
+      "loss/crossentropy": 2.0482982218265535,
+      "loss/hidden": 3.483203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1855379816144705,
+      "step": 13690
+    },
+    {
+      "epoch": 0.3425,
+      "grad_norm": 27.875,
+      "grad_norm_var": 98.41432291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2863,
+      "loss/crossentropy": 2.0204505778849127,
+      "loss/hidden": 3.34609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17318440061062573,
+      "step": 13700
+    },
+    {
+      "epoch": 0.34275,
+      "grad_norm": 28.375,
+      "grad_norm_var": 33.1900390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3409,
+      "loss/crossentropy": 2.0783301174640654,
+      "loss/hidden": 3.38125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19091757852584124,
+      "step": 13710
+    },
+    {
+      "epoch": 0.343,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.7150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3412,
+      "loss/crossentropy": 2.0553587220609186,
+      "loss/hidden": 3.39140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18813807256519793,
+      "step": 13720
+    },
+    {
+      "epoch": 0.34325,
+      "grad_norm": 34.25,
+      "grad_norm_var": 1586.7400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4918,
+      "loss/crossentropy": 2.14101425036788,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1929340995848179,
+      "step": 13730
+    },
+    {
+      "epoch": 0.3435,
+      "grad_norm": 32.5,
+      "grad_norm_var": 4.633072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3162,
+      "loss/crossentropy": 2.0531166955828666,
+      "loss/hidden": 3.411328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18961485847830772,
+      "step": 13740
+    },
+    {
+      "epoch": 0.34375,
+      "grad_norm": 51.5,
+      "grad_norm_var": 29.780989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3981,
+      "loss/crossentropy": 2.146902695298195,
+      "loss/hidden": 3.4203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18639619778841734,
+      "step": 13750
+    },
+    {
+      "epoch": 0.344,
+      "grad_norm": 31.875,
+      "grad_norm_var": 27.141080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3596,
+      "loss/crossentropy": 2.2127513885498047,
+      "loss/hidden": 3.407421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19791291914880277,
+      "step": 13760
+    },
+    {
+      "epoch": 0.34425,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.9809895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3245,
+      "loss/crossentropy": 2.113487794995308,
+      "loss/hidden": 3.396875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18752570655196904,
+      "step": 13770
+    },
+    {
+      "epoch": 0.3445,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.7666666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3737,
+      "loss/crossentropy": 2.192716282606125,
+      "loss/hidden": 3.4296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1977760722860694,
+      "step": 13780
+    },
+    {
+      "epoch": 0.34475,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.4999348958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2308,
+      "loss/crossentropy": 2.1602986216545106,
+      "loss/hidden": 3.2796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1859057329595089,
+      "step": 13790
+    },
+    {
+      "epoch": 0.345,
+      "grad_norm": 45.25,
+      "grad_norm_var": 15.587434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3689,
+      "loss/crossentropy": 2.042529730498791,
+      "loss/hidden": 3.34765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2055663662031293,
+      "step": 13800
+    },
+    {
+      "epoch": 0.34525,
+      "grad_norm": 31.375,
+      "grad_norm_var": 16.111458333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.355,
+      "loss/crossentropy": 2.004579763114452,
+      "loss/hidden": 3.46171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18781258668750525,
+      "step": 13810
+    },
+    {
+      "epoch": 0.3455,
+      "grad_norm": 27.75,
+      "grad_norm_var": 2.9369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2946,
+      "loss/crossentropy": 1.9647000446915626,
+      "loss/hidden": 3.404296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18860411625355483,
+      "step": 13820
+    },
+    {
+      "epoch": 0.34575,
+      "grad_norm": 28.625,
+      "grad_norm_var": 2.4830729166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3643,
+      "loss/crossentropy": 2.1840324103832245,
+      "loss/hidden": 3.321875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18433275502175092,
+      "step": 13830
+    },
+    {
+      "epoch": 0.346,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.061393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.39,
+      "loss/crossentropy": 2.1660002395510674,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18589738458395005,
+      "step": 13840
+    },
+    {
+      "epoch": 0.34625,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.237434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2772,
+      "loss/crossentropy": 2.0304790273308755,
+      "loss/hidden": 3.383984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19276629090309144,
+      "step": 13850
+    },
+    {
+      "epoch": 0.3465,
+      "grad_norm": 27.375,
+      "grad_norm_var": 4.35390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2433,
+      "loss/crossentropy": 2.0913878247141837,
+      "loss/hidden": 3.348828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17961858343333006,
+      "step": 13860
+    },
+    {
+      "epoch": 0.34675,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.9833333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.298,
+      "loss/crossentropy": 1.9085475612431764,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1773257221095264,
+      "step": 13870
+    },
+    {
+      "epoch": 0.347,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.2061848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3026,
+      "loss/crossentropy": 2.134579537808895,
+      "loss/hidden": 3.29375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18533094711601733,
+      "step": 13880
+    },
+    {
+      "epoch": 0.34725,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.74140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2397,
+      "loss/crossentropy": 2.0479501873254775,
+      "loss/hidden": 3.36015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1840794005431235,
+      "step": 13890
+    },
+    {
+      "epoch": 0.3475,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.6400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2447,
+      "loss/crossentropy": 2.0175841793417932,
+      "loss/hidden": 3.29296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17613419592380525,
+      "step": 13900
+    },
+    {
+      "epoch": 0.34775,
+      "grad_norm": 32.5,
+      "grad_norm_var": 5.333072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3573,
+      "loss/crossentropy": 2.248972457647324,
+      "loss/hidden": 3.35234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19207954667508603,
+      "step": 13910
+    },
+    {
+      "epoch": 0.348,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.843489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3649,
+      "loss/crossentropy": 2.215424671769142,
+      "loss/hidden": 3.3328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19513901360332966,
+      "step": 13920
+    },
+    {
+      "epoch": 0.34825,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.004622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4324,
+      "loss/crossentropy": 2.157894307374954,
+      "loss/hidden": 3.457421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20247049070894718,
+      "step": 13930
+    },
+    {
+      "epoch": 0.3485,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.4150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3422,
+      "loss/crossentropy": 2.14641355201602,
+      "loss/hidden": 3.37890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1898975013755262,
+      "step": 13940
+    },
+    {
+      "epoch": 0.34875,
+      "grad_norm": 28.5,
+      "grad_norm_var": 6.0103515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3309,
+      "loss/crossentropy": 2.0432268232107162,
+      "loss/hidden": 3.461328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19791582636535168,
+      "step": 13950
+    },
+    {
+      "epoch": 0.349,
+      "grad_norm": 28.875,
+      "grad_norm_var": 4.4744140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3187,
+      "loss/crossentropy": 2.1648118555545808,
+      "loss/hidden": 3.301171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1794663654640317,
+      "step": 13960
+    },
+    {
+      "epoch": 0.34925,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.1332682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3122,
+      "loss/crossentropy": 2.0566162675619126,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18033871669322252,
+      "step": 13970
+    },
+    {
+      "epoch": 0.3495,
+      "grad_norm": 27.875,
+      "grad_norm_var": 1.3686848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2538,
+      "loss/crossentropy": 2.0309479638934134,
+      "loss/hidden": 3.316015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1834420131519437,
+      "step": 13980
+    },
+    {
+      "epoch": 0.34975,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.3510416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3068,
+      "loss/crossentropy": 2.0843526370823384,
+      "loss/hidden": 3.35859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18843147568404675,
+      "step": 13990
+    },
+    {
+      "epoch": 0.35,
+      "grad_norm": 34.25,
+      "grad_norm_var": 10.67265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3839,
+      "loss/crossentropy": 2.0708370715379716,
+      "loss/hidden": 3.389453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19229107201099396,
+      "step": 14000
+    },
+    {
+      "epoch": 0.35025,
+      "grad_norm": 28.5,
+      "grad_norm_var": 11.178059895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3404,
+      "loss/crossentropy": 2.016050732135773,
+      "loss/hidden": 3.49921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1993856718763709,
+      "step": 14010
+    },
+    {
+      "epoch": 0.3505,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.55625,
+      "learning_rate": 0.0001,
+      "loss": 7.4397,
+      "loss/crossentropy": 2.0941019743680953,
+      "loss/hidden": 3.445703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1941377494484186,
+      "step": 14020
+    },
+    {
+      "epoch": 0.35075,
+      "grad_norm": 28.125,
+      "grad_norm_var": 1.7934895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2787,
+      "loss/crossentropy": 2.175595435500145,
+      "loss/hidden": 3.36484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1906340267509222,
+      "step": 14030
+    },
+    {
+      "epoch": 0.351,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.01015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3322,
+      "loss/crossentropy": 1.945583702623844,
+      "loss/hidden": 3.308203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17433239622041583,
+      "step": 14040
+    },
+    {
+      "epoch": 0.35125,
+      "grad_norm": 32.0,
+      "grad_norm_var": 11.308072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4361,
+      "loss/crossentropy": 2.078453540802002,
+      "loss/hidden": 3.271484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1803632376715541,
+      "step": 14050
+    },
+    {
+      "epoch": 0.3515,
+      "grad_norm": 31.125,
+      "grad_norm_var": 11.920833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3917,
+      "loss/crossentropy": 2.2025377944111826,
+      "loss/hidden": 3.38515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18862273804843427,
+      "step": 14060
+    },
+    {
+      "epoch": 0.35175,
+      "grad_norm": 26.875,
+      "grad_norm_var": 14.82890625,
+      "learning_rate": 0.0001,
+      "loss": 7.294,
+      "loss/crossentropy": 2.0608675971627237,
+      "loss/hidden": 3.229296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17895151115953922,
+      "step": 14070
+    },
+    {
+      "epoch": 0.352,
+      "grad_norm": 31.5,
+      "grad_norm_var": 16.512239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2397,
+      "loss/crossentropy": 1.951252208650112,
+      "loss/hidden": 3.347265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17488121166825293,
+      "step": 14080
+    },
+    {
+      "epoch": 0.35225,
+      "grad_norm": 29.25,
+      "grad_norm_var": 1.6044777159285975e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3104,
+      "loss/crossentropy": 2.1797886729240417,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19085733741521835,
+      "step": 14090
+    },
+    {
+      "epoch": 0.3525,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.6044777150524774e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.269,
+      "loss/crossentropy": 2.133794055879116,
+      "loss/hidden": 3.359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18788108974695206,
+      "step": 14100
+    },
+    {
+      "epoch": 0.35275,
+      "grad_norm": 31.125,
+      "grad_norm_var": 57.800455729166664,
+      "learning_rate": 0.0001,
+      "loss": 7.2859,
+      "loss/crossentropy": 2.0392952769994737,
+      "loss/hidden": 3.317578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1854009686037898,
+      "step": 14110
+    },
+    {
+      "epoch": 0.353,
+      "grad_norm": 29.0,
+      "grad_norm_var": 2.729622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3152,
+      "loss/crossentropy": 2.026089659333229,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18226654808968307,
+      "step": 14120
+    },
+    {
+      "epoch": 0.35325,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.703125,
+      "learning_rate": 0.0001,
+      "loss": 7.4705,
+      "loss/crossentropy": 2.132002358883619,
+      "loss/hidden": 3.39296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2102237056940794,
+      "step": 14130
+    },
+    {
+      "epoch": 0.3535,
+      "grad_norm": 29.0,
+      "grad_norm_var": 2.9497395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3236,
+      "loss/crossentropy": 2.0131492763757706,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19223638353869318,
+      "step": 14140
+    },
+    {
+      "epoch": 0.35375,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.7212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.271,
+      "loss/crossentropy": 2.1130865029990673,
+      "loss/hidden": 3.324609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1914577091112733,
+      "step": 14150
+    },
+    {
+      "epoch": 0.354,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.8254557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4232,
+      "loss/crossentropy": 2.12569759786129,
+      "loss/hidden": 3.460546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19339151214808226,
+      "step": 14160
+    },
+    {
+      "epoch": 0.35425,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1064.6176432291666,
+      "learning_rate": 0.0001,
+      "loss": 7.2083,
+      "loss/crossentropy": 2.0492417976260184,
+      "loss/hidden": 3.37109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18359060864895582,
+      "step": 14170
+    },
+    {
+      "epoch": 0.3545,
+      "grad_norm": 31.375,
+      "grad_norm_var": 9.59765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3908,
+      "loss/crossentropy": 2.099221628904343,
+      "loss/hidden": 3.419140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18481487538665534,
+      "step": 14180
+    },
+    {
+      "epoch": 0.35475,
+      "grad_norm": 44.25,
+      "grad_norm_var": 12.892708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3495,
+      "loss/crossentropy": 2.04138702750206,
+      "loss/hidden": 3.405078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1797080934047699,
+      "step": 14190
+    },
+    {
+      "epoch": 0.355,
+      "grad_norm": 29.875,
+      "grad_norm_var": 15.040559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3724,
+      "loss/crossentropy": 2.0216434367001055,
+      "loss/hidden": 3.440625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19053793977946043,
+      "step": 14200
+    },
+    {
+      "epoch": 0.35525,
+      "grad_norm": 28.25,
+      "grad_norm_var": 13.096809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.299,
+      "loss/crossentropy": 2.160589988529682,
+      "loss/hidden": 3.31953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1827318999916315,
+      "step": 14210
+    },
+    {
+      "epoch": 0.3555,
+      "grad_norm": 31.625,
+      "grad_norm_var": 9.878059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3381,
+      "loss/crossentropy": 1.7803485259413718,
+      "loss/hidden": 3.37890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16204789485782384,
+      "step": 14220
+    },
+    {
+      "epoch": 0.35575,
+      "grad_norm": 28.625,
+      "grad_norm_var": 5.543489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2506,
+      "loss/crossentropy": 1.8810015477240085,
+      "loss/hidden": 3.36640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17873666901141405,
+      "step": 14230
+    },
+    {
+      "epoch": 0.356,
+      "grad_norm": 30.375,
+      "grad_norm_var": 4.208268229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4941,
+      "loss/crossentropy": 2.016931130737066,
+      "loss/hidden": 3.300390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17651305962353944,
+      "step": 14240
+    },
+    {
+      "epoch": 0.35625,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.067122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3659,
+      "loss/crossentropy": 2.1511716455221177,
+      "loss/hidden": 3.334375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18431759886443616,
+      "step": 14250
+    },
+    {
+      "epoch": 0.3565,
+      "grad_norm": 31.5,
+      "grad_norm_var": 14.267708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4284,
+      "loss/crossentropy": 2.0923829920589925,
+      "loss/hidden": 3.414453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20195957068353892,
+      "step": 14260
+    },
+    {
+      "epoch": 0.35675,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.780989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3516,
+      "loss/crossentropy": 2.1890153646469117,
+      "loss/hidden": 3.407421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19206568598747253,
+      "step": 14270
+    },
+    {
+      "epoch": 0.357,
+      "grad_norm": 26.875,
+      "grad_norm_var": 3.0858723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2643,
+      "loss/crossentropy": 2.109933242201805,
+      "loss/hidden": 3.362890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1951975781470537,
+      "step": 14280
+    },
+    {
+      "epoch": 0.35725,
+      "grad_norm": 28.875,
+      "grad_norm_var": 1.7747395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2549,
+      "loss/crossentropy": 2.104505704343319,
+      "loss/hidden": 3.449609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19365676920861005,
+      "step": 14290
+    },
+    {
+      "epoch": 0.3575,
+      "grad_norm": 32.25,
+      "grad_norm_var": 11.114322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.247,
+      "loss/crossentropy": 2.108228546380997,
+      "loss/hidden": 3.398046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18031150791794062,
+      "step": 14300
+    },
+    {
+      "epoch": 0.35775,
+      "grad_norm": 33.5,
+      "grad_norm_var": 4.084375,
+      "learning_rate": 0.0001,
+      "loss": 7.2652,
+      "loss/crossentropy": 2.11306362003088,
+      "loss/hidden": 3.403125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19340378735214472,
+      "step": 14310
+    },
+    {
+      "epoch": 0.358,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.2712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5154,
+      "loss/crossentropy": 2.073298954963684,
+      "loss/hidden": 3.4765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20225481800734996,
+      "step": 14320
+    },
+    {
+      "epoch": 0.35825,
+      "grad_norm": 33.0,
+      "grad_norm_var": 7.609309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3172,
+      "loss/crossentropy": 2.068447032570839,
+      "loss/hidden": 3.360546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1848990023136139,
+      "step": 14330
+    },
+    {
+      "epoch": 0.3585,
+      "grad_norm": 28.625,
+      "grad_norm_var": 9.287434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.341,
+      "loss/crossentropy": 2.0122443050146104,
+      "loss/hidden": 3.270703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16750977858901023,
+      "step": 14340
+    },
+    {
+      "epoch": 0.35875,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.5587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4114,
+      "loss/crossentropy": 1.9815908901393413,
+      "loss/hidden": 3.413671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19162300042808056,
+      "step": 14350
+    },
+    {
+      "epoch": 0.359,
+      "grad_norm": 34.5,
+      "grad_norm_var": 1.75390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4096,
+      "loss/crossentropy": 2.0974107921123504,
+      "loss/hidden": 3.418359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18985395636409522,
+      "step": 14360
+    },
+    {
+      "epoch": 0.35925,
+      "grad_norm": 29.875,
+      "grad_norm_var": 27.2291015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2838,
+      "loss/crossentropy": 2.18302740752697,
+      "loss/hidden": 3.3375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19919114038348198,
+      "step": 14370
+    },
+    {
+      "epoch": 0.3595,
+      "grad_norm": 29.0,
+      "grad_norm_var": 2.393684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3591,
+      "loss/crossentropy": 2.1828642159700395,
+      "loss/hidden": 3.3515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1880602626129985,
+      "step": 14380
+    },
+    {
+      "epoch": 0.35975,
+      "grad_norm": 28.25,
+      "grad_norm_var": 1.8875,
+      "learning_rate": 0.0001,
+      "loss": 7.4149,
+      "loss/crossentropy": 2.0250351071357726,
+      "loss/hidden": 3.432421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21426690481603144,
+      "step": 14390
+    },
+    {
+      "epoch": 0.36,
+      "grad_norm": 36.0,
+      "grad_norm_var": 4.823893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3843,
+      "loss/crossentropy": 1.9927750542759894,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18537398371845484,
+      "step": 14400
+    },
+    {
+      "epoch": 0.36025,
+      "grad_norm": 48.0,
+      "grad_norm_var": 20.270247395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3442,
+      "loss/crossentropy": 2.2051343381404878,
+      "loss/hidden": 3.337109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18966466654092073,
+      "step": 14410
+    },
+    {
+      "epoch": 0.3605,
+      "grad_norm": 26.75,
+      "grad_norm_var": 32.15598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3599,
+      "loss/crossentropy": 2.050907912105322,
+      "loss/hidden": 3.34453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18331205267459155,
+      "step": 14420
+    },
+    {
+      "epoch": 0.36075,
+      "grad_norm": 32.0,
+      "grad_norm_var": 15.5244140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4431,
+      "loss/crossentropy": 1.9177762359380721,
+      "loss/hidden": 3.507421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18733049537986518,
+      "step": 14430
+    },
+    {
+      "epoch": 0.361,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.370247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3748,
+      "loss/crossentropy": 2.1186861246824265,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1846270103007555,
+      "step": 14440
+    },
+    {
+      "epoch": 0.36125,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.378059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4001,
+      "loss/crossentropy": 2.1338623888790607,
+      "loss/hidden": 3.42890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1889139147475362,
+      "step": 14450
+    },
+    {
+      "epoch": 0.3615,
+      "grad_norm": 29.125,
+      "grad_norm_var": 5.045572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3355,
+      "loss/crossentropy": 1.9586334988474845,
+      "loss/hidden": 3.36015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1830888209864497,
+      "step": 14460
+    },
+    {
+      "epoch": 0.36175,
+      "grad_norm": 29.375,
+      "grad_norm_var": 11.797916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4646,
+      "loss/crossentropy": 2.0122861742973326,
+      "loss/hidden": 3.436328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2004035959020257,
+      "step": 14470
+    },
+    {
+      "epoch": 0.362,
+      "grad_norm": 32.0,
+      "grad_norm_var": 11.684309895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4154,
+      "loss/crossentropy": 2.1907495334744453,
+      "loss/hidden": 3.424609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2014760635793209,
+      "step": 14480
+    },
+    {
+      "epoch": 0.36225,
+      "grad_norm": 32.75,
+      "grad_norm_var": 14.495247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3255,
+      "loss/crossentropy": 2.0414177522063257,
+      "loss/hidden": 3.33515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17212325502187015,
+      "step": 14490
+    },
+    {
+      "epoch": 0.3625,
+      "grad_norm": 28.875,
+      "grad_norm_var": 9.85390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4188,
+      "loss/crossentropy": 1.875846453011036,
+      "loss/hidden": 3.4109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18002864755690098,
+      "step": 14500
+    },
+    {
+      "epoch": 0.36275,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.7400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3819,
+      "loss/crossentropy": 2.2282170712947846,
+      "loss/hidden": 3.43359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20617649108171462,
+      "step": 14510
+    },
+    {
+      "epoch": 0.363,
+      "grad_norm": 29.0,
+      "grad_norm_var": 26.622330729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3494,
+      "loss/crossentropy": 2.1207677230238913,
+      "loss/hidden": 3.455859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18537767957895995,
+      "step": 14520
+    },
+    {
+      "epoch": 0.36325,
+      "grad_norm": 28.5,
+      "grad_norm_var": 1.15,
+      "learning_rate": 0.0001,
+      "loss": 7.2906,
+      "loss/crossentropy": 2.1179309889674185,
+      "loss/hidden": 3.39296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18408581465482712,
+      "step": 14530
+    },
+    {
+      "epoch": 0.3635,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.851822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3786,
+      "loss/crossentropy": 2.0486052967607975,
+      "loss/hidden": 3.35859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17906265445053576,
+      "step": 14540
+    },
+    {
+      "epoch": 0.36375,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.184830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3423,
+      "loss/crossentropy": 2.301698251068592,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19077361291274428,
+      "step": 14550
+    },
+    {
+      "epoch": 0.364,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.2134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3047,
+      "loss/crossentropy": 2.074369618296623,
+      "loss/hidden": 3.293359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17446096520870924,
+      "step": 14560
+    },
+    {
+      "epoch": 0.36425,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.6910807291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.412,
+      "loss/crossentropy": 2.070549990236759,
+      "loss/hidden": 3.501953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2114204354584217,
+      "step": 14570
+    },
+    {
+      "epoch": 0.3645,
+      "grad_norm": 31.875,
+      "grad_norm_var": 31.843489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4461,
+      "loss/crossentropy": 2.2021229028701783,
+      "loss/hidden": 3.412890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1946948293596506,
+      "step": 14580
+    },
+    {
+      "epoch": 0.36475,
+      "grad_norm": 28.875,
+      "grad_norm_var": 32.17916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.33,
+      "loss/crossentropy": 2.07574619948864,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.199923849850893,
+      "step": 14590
+    },
+    {
+      "epoch": 0.365,
+      "grad_norm": 29.625,
+      "grad_norm_var": 1.3020833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3285,
+      "loss/crossentropy": 2.236181080341339,
+      "loss/hidden": 3.295703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18559221122413874,
+      "step": 14600
+    },
+    {
+      "epoch": 0.36525,
+      "grad_norm": 27.875,
+      "grad_norm_var": 1.6979166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4122,
+      "loss/crossentropy": 2.029696011543274,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18932779133319855,
+      "step": 14610
+    },
+    {
+      "epoch": 0.3655,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.17265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3296,
+      "loss/crossentropy": 2.0676317110657694,
+      "loss/hidden": 3.308984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1811735849827528,
+      "step": 14620
+    },
+    {
+      "epoch": 0.36575,
+      "grad_norm": 31.5,
+      "grad_norm_var": 3.8629557291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3105,
+      "loss/crossentropy": 2.1635128699243067,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19529270604252816,
+      "step": 14630
+    },
+    {
+      "epoch": 0.366,
+      "grad_norm": 28.25,
+      "grad_norm_var": 3.3955729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3913,
+      "loss/crossentropy": 2.0359733670949938,
+      "loss/hidden": 3.5,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20129154790192844,
+      "step": 14640
+    },
+    {
+      "epoch": 0.36625,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.72265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3418,
+      "loss/crossentropy": 1.895400796085596,
+      "loss/hidden": 3.4125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17521814415231346,
+      "step": 14650
+    },
+    {
+      "epoch": 0.3665,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.7302083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3042,
+      "loss/crossentropy": 2.0394834615290165,
+      "loss/hidden": 3.21953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17233089366927742,
+      "step": 14660
+    },
+    {
+      "epoch": 0.36675,
+      "grad_norm": 28.25,
+      "grad_norm_var": 36.44479166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3335,
+      "loss/crossentropy": 1.8397274687886238,
+      "loss/hidden": 3.4515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17788148634135723,
+      "step": 14670
+    },
+    {
+      "epoch": 0.367,
+      "grad_norm": 27.875,
+      "grad_norm_var": 38.53639322916667,
+      "learning_rate": 0.0001,
+      "loss": 7.3786,
+      "loss/crossentropy": 2.2247194588184356,
+      "loss/hidden": 3.4203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19322688207030297,
+      "step": 14680
+    },
+    {
+      "epoch": 0.36725,
+      "grad_norm": 40.0,
+      "grad_norm_var": 13.45390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3179,
+      "loss/crossentropy": 1.9937927357852459,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17893323805183173,
+      "step": 14690
+    },
+    {
+      "epoch": 0.3675,
+      "grad_norm": 32.75,
+      "grad_norm_var": 11.25,
+      "learning_rate": 0.0001,
+      "loss": 7.2325,
+      "loss/crossentropy": 2.0396489590406417,
+      "loss/hidden": 3.44140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1912191865965724,
+      "step": 14700
+    },
+    {
+      "epoch": 0.36775,
+      "grad_norm": 33.0,
+      "grad_norm_var": 7.471875,
+      "learning_rate": 0.0001,
+      "loss": 7.4509,
+      "loss/crossentropy": 2.1277743950486183,
+      "loss/hidden": 3.449609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20408784411847591,
+      "step": 14710
+    },
+    {
+      "epoch": 0.368,
+      "grad_norm": 30.75,
+      "grad_norm_var": 8.546875,
+      "learning_rate": 0.0001,
+      "loss": 7.3855,
+      "loss/crossentropy": 2.2228190809488297,
+      "loss/hidden": 3.362890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1884877322241664,
+      "step": 14720
+    },
+    {
+      "epoch": 0.36825,
+      "grad_norm": 28.625,
+      "grad_norm_var": 15.142708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4076,
+      "loss/crossentropy": 2.1207469016313554,
+      "loss/hidden": 3.34609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17780339773744344,
+      "step": 14730
+    },
+    {
+      "epoch": 0.3685,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.004622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3209,
+      "loss/crossentropy": 2.0066810354590414,
+      "loss/hidden": 3.519921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20999562088400126,
+      "step": 14740
+    },
+    {
+      "epoch": 0.36875,
+      "grad_norm": 38.25,
+      "grad_norm_var": 15.4259765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4197,
+      "loss/crossentropy": 2.0302488803863525,
+      "loss/hidden": 3.491796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1997592320665717,
+      "step": 14750
+    },
+    {
+      "epoch": 0.369,
+      "grad_norm": 32.5,
+      "grad_norm_var": 22.669791666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3414,
+      "loss/crossentropy": 2.006557123363018,
+      "loss/hidden": 3.333203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18148684445768595,
+      "step": 14760
+    },
+    {
+      "epoch": 0.36925,
+      "grad_norm": 29.5,
+      "grad_norm_var": 22.362434895833335,
+      "learning_rate": 0.0001,
+      "loss": 7.347,
+      "loss/crossentropy": 2.07253782749176,
+      "loss/hidden": 3.274609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17341973297297955,
+      "step": 14770
+    },
+    {
+      "epoch": 0.3695,
+      "grad_norm": 30.75,
+      "grad_norm_var": 17.227018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3839,
+      "loss/crossentropy": 2.056871312856674,
+      "loss/hidden": 3.367578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21048452276736498,
+      "step": 14780
+    },
+    {
+      "epoch": 0.36975,
+      "grad_norm": 27.875,
+      "grad_norm_var": 38.5947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4979,
+      "loss/crossentropy": 2.1910998940467836,
+      "loss/hidden": 3.480859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20320441015064716,
+      "step": 14790
+    },
+    {
+      "epoch": 0.37,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.081788902243793e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3396,
+      "loss/crossentropy": 2.0854848250746727,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19095470141619444,
+      "step": 14800
+    },
+    {
+      "epoch": 0.37025,
+      "grad_norm": 34.75,
+      "grad_norm_var": 2.0817889032778286e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.349,
+      "loss/crossentropy": 2.024295690655708,
+      "loss/hidden": 3.417578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2073504414409399,
+      "step": 14810
+    },
+    {
+      "epoch": 0.3705,
+      "grad_norm": 30.375,
+      "grad_norm_var": 8.116080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.1984,
+      "loss/crossentropy": 1.9738368421792984,
+      "loss/hidden": 3.352734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.175398519821465,
+      "step": 14820
+    },
+    {
+      "epoch": 0.37075,
+      "grad_norm": 34.0,
+      "grad_norm_var": 3.95390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3269,
+      "loss/crossentropy": 2.0109583623707294,
+      "loss/hidden": 3.432421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1902273640036583,
+      "step": 14830
+    },
+    {
+      "epoch": 0.371,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.410416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3309,
+      "loss/crossentropy": 2.0899639569222925,
+      "loss/hidden": 3.444921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18828139845281838,
+      "step": 14840
+    },
+    {
+      "epoch": 0.37125,
+      "grad_norm": 31.375,
+      "grad_norm_var": 51.5087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2761,
+      "loss/crossentropy": 2.0561746567487718,
+      "loss/hidden": 3.308203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17576684867963194,
+      "step": 14850
+    },
+    {
+      "epoch": 0.3715,
+      "grad_norm": 32.25,
+      "grad_norm_var": 54.873372395833336,
+      "learning_rate": 0.0001,
+      "loss": 7.2875,
+      "loss/crossentropy": 2.1231576301157475,
+      "loss/hidden": 3.405859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1990961253643036,
+      "step": 14860
+    },
+    {
+      "epoch": 0.37175,
+      "grad_norm": 34.25,
+      "grad_norm_var": 29.984309895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3424,
+      "loss/crossentropy": 2.1051357120275496,
+      "loss/hidden": 3.29453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17979057859629394,
+      "step": 14870
+    },
+    {
+      "epoch": 0.372,
+      "grad_norm": 28.25,
+      "grad_norm_var": 19850.327018229167,
+      "learning_rate": 0.0001,
+      "loss": 7.3823,
+      "loss/crossentropy": 2.3242007076740263,
+      "loss/hidden": 3.49609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2740444682538509,
+      "step": 14880
+    },
+    {
+      "epoch": 0.37225,
+      "grad_norm": 27.25,
+      "grad_norm_var": 35.0625,
+      "learning_rate": 0.0001,
+      "loss": 7.2273,
+      "loss/crossentropy": 2.143503928184509,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19005000293254853,
+      "step": 14890
+    },
+    {
+      "epoch": 0.3725,
+      "grad_norm": 34.0,
+      "grad_norm_var": 38.23125,
+      "learning_rate": 0.0001,
+      "loss": 7.2655,
+      "loss/crossentropy": 2.1128626547753813,
+      "loss/hidden": 3.40703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18479138296097516,
+      "step": 14900
+    },
+    {
+      "epoch": 0.37275,
+      "grad_norm": 31.0,
+      "grad_norm_var": 26.7478515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3153,
+      "loss/crossentropy": 1.860854334384203,
+      "loss/hidden": 3.255859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1592037882655859,
+      "step": 14910
+    },
+    {
+      "epoch": 0.373,
+      "grad_norm": 29.625,
+      "grad_norm_var": 30.737239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.287,
+      "loss/crossentropy": 2.067621612548828,
+      "loss/hidden": 3.468359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1897974604740739,
+      "step": 14920
+    },
+    {
+      "epoch": 0.37325,
+      "grad_norm": 29.0,
+      "grad_norm_var": 51.01223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4038,
+      "loss/crossentropy": 2.0471679329872132,
+      "loss/hidden": 3.41484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18869173359125851,
+      "step": 14930
+    },
+    {
+      "epoch": 0.3735,
+      "grad_norm": 40.25,
+      "grad_norm_var": 42.828059895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2638,
+      "loss/crossentropy": 2.0969312518835066,
+      "loss/hidden": 3.308984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19103059088811278,
+      "step": 14940
+    },
+    {
+      "epoch": 0.37375,
+      "grad_norm": 27.625,
+      "grad_norm_var": 16.3625,
+      "learning_rate": 0.0001,
+      "loss": 7.3022,
+      "loss/crossentropy": 2.208647185564041,
+      "loss/hidden": 3.323046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1936396975070238,
+      "step": 14950
+    },
+    {
+      "epoch": 0.374,
+      "grad_norm": 28.625,
+      "grad_norm_var": 16.837239583333332,
+      "learning_rate": 0.0001,
+      "loss": 7.2414,
+      "loss/crossentropy": 2.1974997609853744,
+      "loss/hidden": 3.340625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19469811543822288,
+      "step": 14960
+    },
+    {
+      "epoch": 0.37425,
+      "grad_norm": 29.625,
+      "grad_norm_var": 14.108072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2849,
+      "loss/crossentropy": 2.134792809188366,
+      "loss/hidden": 3.20625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17098608147352934,
+      "step": 14970
+    },
+    {
+      "epoch": 0.3745,
+      "grad_norm": 33.25,
+      "grad_norm_var": 27.1181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3014,
+      "loss/crossentropy": 2.1276439666748046,
+      "loss/hidden": 3.3859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18990135621279478,
+      "step": 14980
+    },
+    {
+      "epoch": 0.37475,
+      "grad_norm": 29.75,
+      "grad_norm_var": 16.033072916666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3335,
+      "loss/crossentropy": 2.0856949634850026,
+      "loss/hidden": 3.3265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18325317203998565,
+      "step": 14990
+    },
+    {
+      "epoch": 0.375,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.321875,
+      "learning_rate": 0.0001,
+      "loss": 7.3642,
+      "loss/crossentropy": 2.1136090487241743,
+      "loss/hidden": 3.523828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20829910095781087,
+      "step": 15000
+    },
+    {
+      "epoch": 0.37525,
+      "grad_norm": 33.25,
+      "grad_norm_var": 4.076497395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3349,
+      "loss/crossentropy": 2.007338209450245,
+      "loss/hidden": 3.323828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17481161141768098,
+      "step": 15010
+    },
+    {
+      "epoch": 0.3755,
+      "grad_norm": 30.125,
+      "grad_norm_var": 7.48515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3839,
+      "loss/crossentropy": 2.1611456774175166,
+      "loss/hidden": 3.3484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1868826244957745,
+      "step": 15020
+    },
+    {
+      "epoch": 0.37575,
+      "grad_norm": 30.0,
+      "grad_norm_var": 5.459375,
+      "learning_rate": 0.0001,
+      "loss": 7.2878,
+      "loss/crossentropy": 2.0543340116739275,
+      "loss/hidden": 3.440625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1872962446883321,
+      "step": 15030
+    },
+    {
+      "epoch": 0.376,
+      "grad_norm": 34.5,
+      "grad_norm_var": 4.847916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2701,
+      "loss/crossentropy": 2.079775569587946,
+      "loss/hidden": 3.40546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19110171888023614,
+      "step": 15040
+    },
+    {
+      "epoch": 0.37625,
+      "grad_norm": 29.5,
+      "grad_norm_var": 4.3212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3025,
+      "loss/crossentropy": 2.267373038828373,
+      "loss/hidden": 3.31015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1936658274382353,
+      "step": 15050
+    },
+    {
+      "epoch": 0.3765,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.746875,
+      "learning_rate": 0.0001,
+      "loss": 7.3107,
+      "loss/crossentropy": 2.1074527725577354,
+      "loss/hidden": 3.3203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18790427446365357,
+      "step": 15060
+    },
+    {
+      "epoch": 0.37675,
+      "grad_norm": 29.125,
+      "grad_norm_var": 5.162955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3127,
+      "loss/crossentropy": 2.1202156826853753,
+      "loss/hidden": 3.38671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18663588212803006,
+      "step": 15070
+    },
+    {
+      "epoch": 0.377,
+      "grad_norm": 32.5,
+      "grad_norm_var": 4.164583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.302,
+      "loss/crossentropy": 1.978425008058548,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17640805542469024,
+      "step": 15080
+    },
+    {
+      "epoch": 0.37725,
+      "grad_norm": 29.75,
+      "grad_norm_var": 5.9962890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2746,
+      "loss/crossentropy": 1.9541339822113515,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18182314159348606,
+      "step": 15090
+    },
+    {
+      "epoch": 0.3775,
+      "grad_norm": 31.375,
+      "grad_norm_var": 5.4822265625,
+      "learning_rate": 0.0001,
+      "loss": 7.1614,
+      "loss/crossentropy": 2.26050483584404,
+      "loss/hidden": 3.287109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18672147747129203,
+      "step": 15100
+    },
+    {
+      "epoch": 0.37775,
+      "grad_norm": 35.75,
+      "grad_norm_var": 10.6431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4291,
+      "loss/crossentropy": 1.9852021753787994,
+      "loss/hidden": 3.343359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1927258798852563,
+      "step": 15110
+    },
+    {
+      "epoch": 0.378,
+      "grad_norm": 30.875,
+      "grad_norm_var": 6.601041666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2829,
+      "loss/crossentropy": 2.067719455063343,
+      "loss/hidden": 3.293359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1875240265391767,
+      "step": 15120
+    },
+    {
+      "epoch": 0.37825,
+      "grad_norm": 34.75,
+      "grad_norm_var": 8.2572265625,
+      "learning_rate": 0.0001,
+      "loss": 7.206,
+      "loss/crossentropy": 2.0955534875392914,
+      "loss/hidden": 3.359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18975083716213703,
+      "step": 15130
+    },
+    {
+      "epoch": 0.3785,
+      "grad_norm": 33.5,
+      "grad_norm_var": 4.514518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3649,
+      "loss/crossentropy": 2.231545624136925,
+      "loss/hidden": 3.4015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1933881167322397,
+      "step": 15140
+    },
+    {
+      "epoch": 0.37875,
+      "grad_norm": 31.25,
+      "grad_norm_var": 6.922916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3186,
+      "loss/crossentropy": 2.1041683927178383,
+      "loss/hidden": 3.315625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17568105049431323,
+      "step": 15150
+    },
+    {
+      "epoch": 0.379,
+      "grad_norm": 31.25,
+      "grad_norm_var": 8.064583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1475,
+      "loss/crossentropy": 2.128729820251465,
+      "loss/hidden": 3.267578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18820853671059012,
+      "step": 15160
+    },
+    {
+      "epoch": 0.37925,
+      "grad_norm": 32.75,
+      "grad_norm_var": 54.0400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4073,
+      "loss/crossentropy": 2.0806841500103475,
+      "loss/hidden": 3.43515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19645702093839645,
+      "step": 15170
+    },
+    {
+      "epoch": 0.3795,
+      "grad_norm": 31.375,
+      "grad_norm_var": 53.652018229166664,
+      "learning_rate": 0.0001,
+      "loss": 7.1232,
+      "loss/crossentropy": 2.077859415113926,
+      "loss/hidden": 3.314453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18113754820078612,
+      "step": 15180
+    },
+    {
+      "epoch": 0.37975,
+      "grad_norm": 34.0,
+      "grad_norm_var": 2.986167996076768e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2362,
+      "loss/crossentropy": 1.911629044264555,
+      "loss/hidden": 3.439453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17613004967570306,
+      "step": 15190
+    },
+    {
+      "epoch": 0.38,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.9861679996912794e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2782,
+      "loss/crossentropy": 2.120191100984812,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1851762106642127,
+      "step": 15200
+    },
+    {
+      "epoch": 0.38025,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.6010416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3305,
+      "loss/crossentropy": 2.084269215166569,
+      "loss/hidden": 3.39453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1797555785626173,
+      "step": 15210
+    },
+    {
+      "epoch": 0.3805,
+      "grad_norm": 31.625,
+      "grad_norm_var": 10.604622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2582,
+      "loss/crossentropy": 2.077724662423134,
+      "loss/hidden": 3.292578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17972503434866666,
+      "step": 15220
+    },
+    {
+      "epoch": 0.38075,
+      "grad_norm": 32.25,
+      "grad_norm_var": 10.440625,
+      "learning_rate": 0.0001,
+      "loss": 7.298,
+      "loss/crossentropy": 2.025072456151247,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18922405112534763,
+      "step": 15230
+    },
+    {
+      "epoch": 0.381,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.56875,
+      "learning_rate": 0.0001,
+      "loss": 7.3076,
+      "loss/crossentropy": 2.1376619666814802,
+      "loss/hidden": 3.40078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2057780459523201,
+      "step": 15240
+    },
+    {
+      "epoch": 0.38125,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.3330729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2858,
+      "loss/crossentropy": 2.0143167987465858,
+      "loss/hidden": 3.396875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19064239636063576,
+      "step": 15250
+    },
+    {
+      "epoch": 0.3815,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.2483723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2522,
+      "loss/crossentropy": 2.0520439468324185,
+      "loss/hidden": 3.332421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18599720411002635,
+      "step": 15260
+    },
+    {
+      "epoch": 0.38175,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.7712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.1878,
+      "loss/crossentropy": 2.1704175978899003,
+      "loss/hidden": 3.32578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17961867786943914,
+      "step": 15270
+    },
+    {
+      "epoch": 0.382,
+      "grad_norm": 30.375,
+      "grad_norm_var": 7.103125,
+      "learning_rate": 0.0001,
+      "loss": 7.3041,
+      "loss/crossentropy": 2.081410513818264,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18557187523692847,
+      "step": 15280
+    },
+    {
+      "epoch": 0.38225,
+      "grad_norm": 32.25,
+      "grad_norm_var": 6.169205729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2756,
+      "loss/crossentropy": 2.0742292061448095,
+      "loss/hidden": 3.355078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18702635429799558,
+      "step": 15290
+    },
+    {
+      "epoch": 0.3825,
+      "grad_norm": 29.125,
+      "grad_norm_var": 5.061458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3008,
+      "loss/crossentropy": 2.081205701082945,
+      "loss/hidden": 3.38828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1858987484127283,
+      "step": 15300
+    },
+    {
+      "epoch": 0.38275,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.6405598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2722,
+      "loss/crossentropy": 1.9639541752636434,
+      "loss/hidden": 3.383203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18367136251181365,
+      "step": 15310
+    },
+    {
+      "epoch": 0.383,
+      "grad_norm": 29.375,
+      "grad_norm_var": 4.342708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3282,
+      "loss/crossentropy": 1.8718928024172783,
+      "loss/hidden": 3.343359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18070896286517382,
+      "step": 15320
+    },
+    {
+      "epoch": 0.38325,
+      "grad_norm": 29.75,
+      "grad_norm_var": 15.129166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.1724,
+      "loss/crossentropy": 2.0988180577754973,
+      "loss/hidden": 3.3046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1885043691843748,
+      "step": 15330
+    },
+    {
+      "epoch": 0.3835,
+      "grad_norm": 28.5,
+      "grad_norm_var": 11.845833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.385,
+      "loss/crossentropy": 2.1939360052347183,
+      "loss/hidden": 3.377734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1921105194836855,
+      "step": 15340
+    },
+    {
+      "epoch": 0.38375,
+      "grad_norm": 37.0,
+      "grad_norm_var": 9.03515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3133,
+      "loss/crossentropy": 2.0929661616683006,
+      "loss/hidden": 3.3015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17670788932591677,
+      "step": 15350
+    },
+    {
+      "epoch": 0.384,
+      "grad_norm": 30.625,
+      "grad_norm_var": 6.0259765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2666,
+      "loss/crossentropy": 2.146226355433464,
+      "loss/hidden": 3.27578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18036798052489758,
+      "step": 15360
+    },
+    {
+      "epoch": 0.38425,
+      "grad_norm": 33.75,
+      "grad_norm_var": 6.095768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2881,
+      "loss/crossentropy": 2.1441670656204224,
+      "loss/hidden": 3.29609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17735737133771182,
+      "step": 15370
+    },
+    {
+      "epoch": 0.3845,
+      "grad_norm": 31.0,
+      "grad_norm_var": 15.9291015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2613,
+      "loss/crossentropy": 1.9772998243570328,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1872927848249674,
+      "step": 15380
+    },
+    {
+      "epoch": 0.38475,
+      "grad_norm": 29.125,
+      "grad_norm_var": 23.395833333333332,
+      "learning_rate": 0.0001,
+      "loss": 7.3075,
+      "loss/crossentropy": 2.061436576396227,
+      "loss/hidden": 3.303515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1824639480561018,
+      "step": 15390
+    },
+    {
+      "epoch": 0.385,
+      "grad_norm": 28.0,
+      "grad_norm_var": 40.42890625,
+      "learning_rate": 0.0001,
+      "loss": 7.307,
+      "loss/crossentropy": 2.222813582420349,
+      "loss/hidden": 3.35078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1908612387254834,
+      "step": 15400
+    },
+    {
+      "epoch": 0.38525,
+      "grad_norm": 33.0,
+      "grad_norm_var": 6.859309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3366,
+      "loss/crossentropy": 1.905498656630516,
+      "loss/hidden": 3.4296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19282873664051295,
+      "step": 15410
+    },
+    {
+      "epoch": 0.3855,
+      "grad_norm": 34.0,
+      "grad_norm_var": 2.953125,
+      "learning_rate": 0.0001,
+      "loss": 7.3019,
+      "loss/crossentropy": 2.1414296105504036,
+      "loss/hidden": 3.362109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1922605263069272,
+      "step": 15420
+    },
+    {
+      "epoch": 0.38575,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.1556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3754,
+      "loss/crossentropy": 2.168923759460449,
+      "loss/hidden": 3.37890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18643623348325492,
+      "step": 15430
+    },
+    {
+      "epoch": 0.386,
+      "grad_norm": 47.5,
+      "grad_norm_var": 21.223893229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.311,
+      "loss/crossentropy": 2.0150568321347238,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18107260223478078,
+      "step": 15440
+    },
+    {
+      "epoch": 0.38625,
+      "grad_norm": 30.125,
+      "grad_norm_var": 22.613541666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3546,
+      "loss/crossentropy": 2.000614331662655,
+      "loss/hidden": 3.412890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1795403585769236,
+      "step": 15450
+    },
+    {
+      "epoch": 0.3865,
+      "grad_norm": 30.875,
+      "grad_norm_var": 4.323958333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2787,
+      "loss/crossentropy": 2.1310448944568634,
+      "loss/hidden": 3.3625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20316302124410868,
+      "step": 15460
+    },
+    {
+      "epoch": 0.38675,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.8900390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2955,
+      "loss/crossentropy": 2.261080901324749,
+      "loss/hidden": 3.283203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.195007973536849,
+      "step": 15470
+    },
+    {
+      "epoch": 0.387,
+      "grad_norm": 34.5,
+      "grad_norm_var": 5.880208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.412,
+      "loss/crossentropy": 2.137118950486183,
+      "loss/hidden": 3.428125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20062100179493428,
+      "step": 15480
+    },
+    {
+      "epoch": 0.38725,
+      "grad_norm": 31.5,
+      "grad_norm_var": 4.9634765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3413,
+      "loss/crossentropy": 2.281751687824726,
+      "loss/hidden": 3.39296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19415688924491406,
+      "step": 15490
+    },
+    {
+      "epoch": 0.3875,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.1134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3408,
+      "loss/crossentropy": 2.0828112423419953,
+      "loss/hidden": 3.463671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20367840733379125,
+      "step": 15500
+    },
+    {
+      "epoch": 0.38775,
+      "grad_norm": 28.75,
+      "grad_norm_var": 67.39765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3114,
+      "loss/crossentropy": 1.944053754210472,
+      "loss/hidden": 3.2671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16787241958081722,
+      "step": 15510
+    },
+    {
+      "epoch": 0.388,
+      "grad_norm": 30.0,
+      "grad_norm_var": 69.334375,
+      "learning_rate": 0.0001,
+      "loss": 7.418,
+      "loss/crossentropy": 2.0586235135793687,
+      "loss/hidden": 3.3078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18477604109793902,
+      "step": 15520
+    },
+    {
+      "epoch": 0.38825,
+      "grad_norm": 30.25,
+      "grad_norm_var": 23.683333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2656,
+      "loss/crossentropy": 2.1504869014024734,
+      "loss/hidden": 3.31953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1810676084831357,
+      "step": 15530
+    },
+    {
+      "epoch": 0.3885,
+      "grad_norm": 30.0,
+      "grad_norm_var": 31.371875,
+      "learning_rate": 0.0001,
+      "loss": 7.2194,
+      "loss/crossentropy": 2.043111228942871,
+      "loss/hidden": 3.349609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18173537645488977,
+      "step": 15540
+    },
+    {
+      "epoch": 0.38875,
+      "grad_norm": 27.75,
+      "grad_norm_var": 25.820247395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2142,
+      "loss/crossentropy": 1.9861251831054687,
+      "loss/hidden": 3.353125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1864440616220236,
+      "step": 15550
+    },
+    {
+      "epoch": 0.389,
+      "grad_norm": 28.875,
+      "grad_norm_var": 31.534830729166668,
+      "learning_rate": 0.0001,
+      "loss": 7.303,
+      "loss/crossentropy": 2.0181840524077415,
+      "loss/hidden": 3.430078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1844530213624239,
+      "step": 15560
+    },
+    {
+      "epoch": 0.38925,
+      "grad_norm": 33.0,
+      "grad_norm_var": 29.174739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3188,
+      "loss/crossentropy": 2.0922836802899836,
+      "loss/hidden": 3.468359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18936851639300584,
+      "step": 15570
+    },
+    {
+      "epoch": 0.3895,
+      "grad_norm": 32.25,
+      "grad_norm_var": 13.853125,
+      "learning_rate": 0.0001,
+      "loss": 7.2795,
+      "loss/crossentropy": 2.0901282012462614,
+      "loss/hidden": 3.361328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18346477318555116,
+      "step": 15580
+    },
+    {
+      "epoch": 0.38975,
+      "grad_norm": 50.0,
+      "grad_norm_var": 32.6509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2721,
+      "loss/crossentropy": 2.1004329577088354,
+      "loss/hidden": 3.362890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1858608016744256,
+      "step": 15590
+    },
+    {
+      "epoch": 0.39,
+      "grad_norm": 28.5,
+      "grad_norm_var": 60.52389322916667,
+      "learning_rate": 0.0001,
+      "loss": 7.1135,
+      "loss/crossentropy": 2.058871814608574,
+      "loss/hidden": 3.345703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18565063830465078,
+      "step": 15600
+    },
+    {
+      "epoch": 0.39025,
+      "grad_norm": 32.25,
+      "grad_norm_var": 45.546875,
+      "learning_rate": 0.0001,
+      "loss": 7.1575,
+      "loss/crossentropy": 1.9910513393580913,
+      "loss/hidden": 3.268359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17116884887218475,
+      "step": 15610
+    },
+    {
+      "epoch": 0.3905,
+      "grad_norm": 28.625,
+      "grad_norm_var": 13.693489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1938,
+      "loss/crossentropy": 2.0928867653012277,
+      "loss/hidden": 3.313671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1812552411109209,
+      "step": 15620
+    },
+    {
+      "epoch": 0.39075,
+      "grad_norm": 38.0,
+      "grad_norm_var": 17.55390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2295,
+      "loss/crossentropy": 1.908367295563221,
+      "loss/hidden": 3.356640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17530275397002698,
+      "step": 15630
+    },
+    {
+      "epoch": 0.391,
+      "grad_norm": 32.5,
+      "grad_norm_var": 16.020572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3788,
+      "loss/crossentropy": 2.1054799772799013,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18641111236065627,
+      "step": 15640
+    },
+    {
+      "epoch": 0.39125,
+      "grad_norm": 37.5,
+      "grad_norm_var": 17.2744140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2224,
+      "loss/crossentropy": 2.0865116521716116,
+      "loss/hidden": 3.315234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1855388769879937,
+      "step": 15650
+    },
+    {
+      "epoch": 0.3915,
+      "grad_norm": 30.5,
+      "grad_norm_var": 10.327018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3264,
+      "loss/crossentropy": 2.0367104552686213,
+      "loss/hidden": 3.38828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18481685388833285,
+      "step": 15660
+    },
+    {
+      "epoch": 0.39175,
+      "grad_norm": 35.25,
+      "grad_norm_var": 8.862239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.343,
+      "loss/crossentropy": 2.2644485861063,
+      "loss/hidden": 3.343359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20473828464746474,
+      "step": 15670
+    },
+    {
+      "epoch": 0.392,
+      "grad_norm": 29.625,
+      "grad_norm_var": 9.986458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3652,
+      "loss/crossentropy": 2.039817491173744,
+      "loss/hidden": 3.337890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1852146415039897,
+      "step": 15680
+    },
+    {
+      "epoch": 0.39225,
+      "grad_norm": 36.5,
+      "grad_norm_var": 11.512239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2599,
+      "loss/crossentropy": 2.0308106377720834,
+      "loss/hidden": 3.42421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19251971375197172,
+      "step": 15690
+    },
+    {
+      "epoch": 0.3925,
+      "grad_norm": 30.375,
+      "grad_norm_var": 9.9650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2512,
+      "loss/crossentropy": 1.980294729024172,
+      "loss/hidden": 3.2890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17508508097380399,
+      "step": 15700
+    },
+    {
+      "epoch": 0.39275,
+      "grad_norm": 32.25,
+      "grad_norm_var": 8.9869140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2101,
+      "loss/crossentropy": 1.817311129719019,
+      "loss/hidden": 3.31328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17575446693226696,
+      "step": 15710
+    },
+    {
+      "epoch": 0.393,
+      "grad_norm": 32.0,
+      "grad_norm_var": 6.434375,
+      "learning_rate": 0.0001,
+      "loss": 7.3109,
+      "loss/crossentropy": 2.231336107850075,
+      "loss/hidden": 3.253515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.186120249889791,
+      "step": 15720
+    },
+    {
+      "epoch": 0.39325,
+      "grad_norm": 27.5,
+      "grad_norm_var": 6.1375,
+      "learning_rate": 0.0001,
+      "loss": 7.2341,
+      "loss/crossentropy": 2.1372755438089373,
+      "loss/hidden": 3.44140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19311708956956863,
+      "step": 15730
+    },
+    {
+      "epoch": 0.3935,
+      "grad_norm": 33.0,
+      "grad_norm_var": 12.706184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4575,
+      "loss/crossentropy": 2.1330266147851944,
+      "loss/hidden": 3.3796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19104793900623918,
+      "step": 15740
+    },
+    {
+      "epoch": 0.39375,
+      "grad_norm": 29.625,
+      "grad_norm_var": 14.846875,
+      "learning_rate": 0.0001,
+      "loss": 7.3186,
+      "loss/crossentropy": 2.090308539569378,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20190660767257212,
+      "step": 15750
+    },
+    {
+      "epoch": 0.394,
+      "grad_norm": 30.0,
+      "grad_norm_var": 7.740559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3049,
+      "loss/crossentropy": 2.075870428979397,
+      "loss/hidden": 3.344140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1769866080954671,
+      "step": 15760
+    },
+    {
+      "epoch": 0.39425,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.682291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2318,
+      "loss/crossentropy": 2.072967618703842,
+      "loss/hidden": 3.380859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18285961151123048,
+      "step": 15770
+    },
+    {
+      "epoch": 0.3945,
+      "grad_norm": 28.0,
+      "grad_norm_var": 16.6494140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2528,
+      "loss/crossentropy": 2.099040774255991,
+      "loss/hidden": 3.35546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18526200093328954,
+      "step": 15780
+    },
+    {
+      "epoch": 0.39475,
+      "grad_norm": 30.75,
+      "grad_norm_var": 13.454622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2255,
+      "loss/crossentropy": 2.0226017188280823,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17934049041941763,
+      "step": 15790
+    },
+    {
+      "epoch": 0.395,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.4546223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3046,
+      "loss/crossentropy": 2.1888489231467245,
+      "loss/hidden": 3.294921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18036039546132088,
+      "step": 15800
+    },
+    {
+      "epoch": 0.39525,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.6393229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2127,
+      "loss/crossentropy": 1.9043663069605827,
+      "loss/hidden": 3.309375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17555271480232476,
+      "step": 15810
+    },
+    {
+      "epoch": 0.3955,
+      "grad_norm": 31.75,
+      "grad_norm_var": 85.26223958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2355,
+      "loss/crossentropy": 1.9372950181365014,
+      "loss/hidden": 3.383203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17982141822576522,
+      "step": 15820
+    },
+    {
+      "epoch": 0.39575,
+      "grad_norm": 30.375,
+      "grad_norm_var": 86.18795572916666,
+      "learning_rate": 0.0001,
+      "loss": 7.2905,
+      "loss/crossentropy": 2.2129892081022264,
+      "loss/hidden": 3.3265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19481155294924973,
+      "step": 15830
+    },
+    {
+      "epoch": 0.396,
+      "grad_norm": 31.75,
+      "grad_norm_var": 4.287434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1819,
+      "loss/crossentropy": 1.9988259918987752,
+      "loss/hidden": 3.203515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16147571448236703,
+      "step": 15840
+    },
+    {
+      "epoch": 0.39625,
+      "grad_norm": 28.875,
+      "grad_norm_var": 4.6009765625,
+      "learning_rate": 0.0001,
+      "loss": 7.359,
+      "loss/crossentropy": 2.1142133861780166,
+      "loss/hidden": 3.320703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18781689032912255,
+      "step": 15850
+    },
+    {
+      "epoch": 0.3965,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.5452473958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.327,
+      "loss/crossentropy": 2.0223607540130617,
+      "loss/hidden": 3.265234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19158173706382514,
+      "step": 15860
+    },
+    {
+      "epoch": 0.39675,
+      "grad_norm": 28.625,
+      "grad_norm_var": 1.7729166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3644,
+      "loss/crossentropy": 2.170432631671429,
+      "loss/hidden": 3.266015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18453630786389114,
+      "step": 15870
+    },
+    {
+      "epoch": 0.397,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.5559895833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3114,
+      "loss/crossentropy": 1.9615431509912014,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18825806053355337,
+      "step": 15880
+    },
+    {
+      "epoch": 0.39725,
+      "grad_norm": 37.25,
+      "grad_norm_var": 71.86555989583333,
+      "learning_rate": 0.0001,
+      "loss": 7.3443,
+      "loss/crossentropy": 2.0923714101314546,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18501388970762492,
+      "step": 15890
+    },
+    {
+      "epoch": 0.3975,
+      "grad_norm": 29.25,
+      "grad_norm_var": 75.43307291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2744,
+      "loss/crossentropy": 2.0929214730858803,
+      "loss/hidden": 3.30390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17970609571784735,
+      "step": 15900
+    },
+    {
+      "epoch": 0.39775,
+      "grad_norm": 31.375,
+      "grad_norm_var": 4.829622395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2885,
+      "loss/crossentropy": 1.9244549985975028,
+      "loss/hidden": 3.2625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16870631305500866,
+      "step": 15910
+    },
+    {
+      "epoch": 0.398,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.684309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2604,
+      "loss/crossentropy": 2.024969767779112,
+      "loss/hidden": 3.30625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17861165013164282,
+      "step": 15920
+    },
+    {
+      "epoch": 0.39825,
+      "grad_norm": 36.5,
+      "grad_norm_var": 13.322330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2294,
+      "loss/crossentropy": 1.9006286635994911,
+      "loss/hidden": 3.35703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1897978626191616,
+      "step": 15930
+    },
+    {
+      "epoch": 0.3985,
+      "grad_norm": 41.25,
+      "grad_norm_var": 19.4478515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3494,
+      "loss/crossentropy": 2.1528166741132737,
+      "loss/hidden": 3.35546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20025787688791752,
+      "step": 15940
+    },
+    {
+      "epoch": 0.39875,
+      "grad_norm": 30.125,
+      "grad_norm_var": 12.692122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2807,
+      "loss/crossentropy": 2.159406080842018,
+      "loss/hidden": 3.341015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19357968419790267,
+      "step": 15950
+    },
+    {
+      "epoch": 0.399,
+      "grad_norm": 28.375,
+      "grad_norm_var": 9.457291666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.0891,
+      "loss/crossentropy": 2.0662672072649,
+      "loss/hidden": 3.35625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1739374803379178,
+      "step": 15960
+    },
+    {
+      "epoch": 0.39925,
+      "grad_norm": 33.25,
+      "grad_norm_var": 17.795833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3368,
+      "loss/crossentropy": 2.2619366616010668,
+      "loss/hidden": 3.3203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19278527814894914,
+      "step": 15970
+    },
+    {
+      "epoch": 0.3995,
+      "grad_norm": 33.5,
+      "grad_norm_var": 32.524739583333336,
+      "learning_rate": 0.0001,
+      "loss": 7.2702,
+      "loss/crossentropy": 1.9777626052498818,
+      "loss/hidden": 3.27109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1701345544308424,
+      "step": 15980
+    },
+    {
+      "epoch": 0.39975,
+      "grad_norm": 31.0,
+      "grad_norm_var": 27.562955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2636,
+      "loss/crossentropy": 1.984829319268465,
+      "loss/hidden": 3.278125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1783665182068944,
+      "step": 15990
+    },
+    {
+      "epoch": 0.4,
+      "grad_norm": 30.25,
+      "grad_norm_var": 7.468489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3522,
+      "loss/crossentropy": 2.027949205040932,
+      "loss/hidden": 3.41328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20451073683798313,
+      "step": 16000
+    },
+    {
+      "epoch": 0.40025,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.1311848958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3003,
+      "loss/crossentropy": 2.0037057876586912,
+      "loss/hidden": 3.309375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1825689303688705,
+      "step": 16010
+    },
+    {
+      "epoch": 0.4005,
+      "grad_norm": 28.625,
+      "grad_norm_var": 4.043684895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3601,
+      "loss/crossentropy": 2.20539278537035,
+      "loss/hidden": 3.33828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19239765033125877,
+      "step": 16020
+    },
+    {
+      "epoch": 0.40075,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.1626528094295864e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2581,
+      "loss/crossentropy": 1.9106431234627963,
+      "loss/hidden": 3.466796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.15896897157654166,
+      "step": 16030
+    },
+    {
+      "epoch": 0.401,
+      "grad_norm": 31.125,
+      "grad_norm_var": 9.828125,
+      "learning_rate": 0.0001,
+      "loss": 7.2905,
+      "loss/crossentropy": 1.9915092147886753,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1976497994735837,
+      "step": 16040
+    },
+    {
+      "epoch": 0.40125,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.596875,
+      "learning_rate": 0.0001,
+      "loss": 7.2439,
+      "loss/crossentropy": 2.0657053992152212,
+      "loss/hidden": 3.266015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17077930886298417,
+      "step": 16050
+    },
+    {
+      "epoch": 0.4015,
+      "grad_norm": 28.625,
+      "grad_norm_var": 7.05390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2314,
+      "loss/crossentropy": 1.9348336219787599,
+      "loss/hidden": 3.45390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18747655171900987,
+      "step": 16060
+    },
+    {
+      "epoch": 0.40175,
+      "grad_norm": 33.0,
+      "grad_norm_var": 10.48515625,
+      "learning_rate": 0.0001,
+      "loss": 7.406,
+      "loss/crossentropy": 2.0627846613526346,
+      "loss/hidden": 3.4453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20208690427243708,
+      "step": 16070
+    },
+    {
+      "epoch": 0.402,
+      "grad_norm": 33.25,
+      "grad_norm_var": 16.162434895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.3501,
+      "loss/crossentropy": 1.984161777049303,
+      "loss/hidden": 3.454296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1927208350971341,
+      "step": 16080
+    },
+    {
+      "epoch": 0.40225,
+      "grad_norm": 32.25,
+      "grad_norm_var": 17.923958333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3793,
+      "loss/crossentropy": 2.074259965121746,
+      "loss/hidden": 3.381640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18814291208982467,
+      "step": 16090
+    },
+    {
+      "epoch": 0.4025,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.3827473958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3529,
+      "loss/crossentropy": 2.1366620868444444,
+      "loss/hidden": 3.35,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1963860308751464,
+      "step": 16100
+    },
+    {
+      "epoch": 0.40275,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.5254557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2352,
+      "loss/crossentropy": 1.9054784037172794,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18159330673515797,
+      "step": 16110
+    },
+    {
+      "epoch": 0.403,
+      "grad_norm": 32.5,
+      "grad_norm_var": 29.687434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4832,
+      "loss/crossentropy": 2.10203175842762,
+      "loss/hidden": 3.530859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21614400669932365,
+      "step": 16120
+    },
+    {
+      "epoch": 0.40325,
+      "grad_norm": 31.25,
+      "grad_norm_var": 29.977018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.1964,
+      "loss/crossentropy": 2.079006028920412,
+      "loss/hidden": 3.35078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1874436529353261,
+      "step": 16130
+    },
+    {
+      "epoch": 0.4035,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.064583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2316,
+      "loss/crossentropy": 2.0732211872935293,
+      "loss/hidden": 3.396875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1887400571256876,
+      "step": 16140
+    },
+    {
+      "epoch": 0.40375,
+      "grad_norm": 31.875,
+      "grad_norm_var": 6.433072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2228,
+      "loss/crossentropy": 1.862874235212803,
+      "loss/hidden": 3.396484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17755231373012065,
+      "step": 16150
+    },
+    {
+      "epoch": 0.404,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.0809895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.307,
+      "loss/crossentropy": 2.0635001718997956,
+      "loss/hidden": 3.3296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17978991996496915,
+      "step": 16160
+    },
+    {
+      "epoch": 0.40425,
+      "grad_norm": 30.375,
+      "grad_norm_var": 13.473958333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2643,
+      "loss/crossentropy": 1.9896802924573422,
+      "loss/hidden": 3.23046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16587578747421503,
+      "step": 16170
+    },
+    {
+      "epoch": 0.4045,
+      "grad_norm": 33.75,
+      "grad_norm_var": 8.254622395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4306,
+      "loss/crossentropy": 2.244751125574112,
+      "loss/hidden": 3.3109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1836428429931402,
+      "step": 16180
+    },
+    {
+      "epoch": 0.40475,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.434477048521451e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.318,
+      "loss/crossentropy": 1.8701960302889347,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18736766083166004,
+      "step": 16190
+    },
+    {
+      "epoch": 0.405,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.434477048703484e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3965,
+      "loss/crossentropy": 2.3353544265031814,
+      "loss/hidden": 3.465625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20016160774976016,
+      "step": 16200
+    },
+    {
+      "epoch": 0.40525,
+      "grad_norm": 28.375,
+      "grad_norm_var": 6.144205729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2603,
+      "loss/crossentropy": 2.022757910192013,
+      "loss/hidden": 3.4578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1913651939481497,
+      "step": 16210
+    },
+    {
+      "epoch": 0.4055,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.872916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2299,
+      "loss/crossentropy": 2.0279084965586662,
+      "loss/hidden": 3.439453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1943587277084589,
+      "step": 16220
+    },
+    {
+      "epoch": 0.40575,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.814322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3063,
+      "loss/crossentropy": 1.9705725610256195,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18895087838172914,
+      "step": 16230
+    },
+    {
+      "epoch": 0.406,
+      "grad_norm": 30.5,
+      "grad_norm_var": 4.430208333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3762,
+      "loss/crossentropy": 2.0867358654737473,
+      "loss/hidden": 3.358984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19180330783128738,
+      "step": 16240
+    },
+    {
+      "epoch": 0.40625,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.9479166666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3468,
+      "loss/crossentropy": 1.9416721653193236,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18548533669672906,
+      "step": 16250
+    },
+    {
+      "epoch": 0.4065,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.3525390625,
+      "learning_rate": 0.0001,
+      "loss": 7.1515,
+      "loss/crossentropy": 1.8505723856389522,
+      "loss/hidden": 3.32265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1714062148705125,
+      "step": 16260
+    },
+    {
+      "epoch": 0.40675,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.9811848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.408,
+      "loss/crossentropy": 2.1564977198839186,
+      "loss/hidden": 3.3796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1916896834038198,
+      "step": 16270
+    },
+    {
+      "epoch": 0.407,
+      "grad_norm": 28.875,
+      "grad_norm_var": 24.1884765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2915,
+      "loss/crossentropy": 2.106581533700228,
+      "loss/hidden": 3.444921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20620792759582401,
+      "step": 16280
+    },
+    {
+      "epoch": 0.40725,
+      "grad_norm": 32.5,
+      "grad_norm_var": 28.264322916666668,
+      "learning_rate": 0.0001,
+      "loss": 7.2376,
+      "loss/crossentropy": 1.9940311327576636,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18459046920761465,
+      "step": 16290
+    },
+    {
+      "epoch": 0.4075,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.088997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2199,
+      "loss/crossentropy": 1.866896914690733,
+      "loss/hidden": 3.371875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17536403173580767,
+      "step": 16300
+    },
+    {
+      "epoch": 0.40775,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.9114583333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3328,
+      "loss/crossentropy": 2.12556598931551,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1877920225262642,
+      "step": 16310
+    },
+    {
+      "epoch": 0.408,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.1035807291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3096,
+      "loss/crossentropy": 1.9810500107705593,
+      "loss/hidden": 3.475390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18928063567727804,
+      "step": 16320
+    },
+    {
+      "epoch": 0.40825,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.1254557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5025,
+      "loss/crossentropy": 2.2035975247621535,
+      "loss/hidden": 3.421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.199692689999938,
+      "step": 16330
+    },
+    {
+      "epoch": 0.4085,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.981184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2389,
+      "loss/crossentropy": 2.0442154854536057,
+      "loss/hidden": 3.263671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17368101943284273,
+      "step": 16340
+    },
+    {
+      "epoch": 0.40875,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.8572265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3269,
+      "loss/crossentropy": 2.0223391756415365,
+      "loss/hidden": 3.396875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19969959184527397,
+      "step": 16350
+    },
+    {
+      "epoch": 0.409,
+      "grad_norm": 31.625,
+      "grad_norm_var": 15.11640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3063,
+      "loss/crossentropy": 1.8919450908899307,
+      "loss/hidden": 3.34921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16754893381148578,
+      "step": 16360
+    },
+    {
+      "epoch": 0.40925,
+      "grad_norm": 32.25,
+      "grad_norm_var": 10.230989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2766,
+      "loss/crossentropy": 2.0608230382204056,
+      "loss/hidden": 3.45859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19851373862475158,
+      "step": 16370
+    },
+    {
+      "epoch": 0.4095,
+      "grad_norm": 30.375,
+      "grad_norm_var": 11.682747395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2895,
+      "loss/crossentropy": 2.0847655333578587,
+      "loss/hidden": 3.3234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.186351899523288,
+      "step": 16380
+    },
+    {
+      "epoch": 0.40975,
+      "grad_norm": 37.75,
+      "grad_norm_var": 14.268684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2556,
+      "loss/crossentropy": 2.010029584169388,
+      "loss/hidden": 3.302734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17632390139624476,
+      "step": 16390
+    },
+    {
+      "epoch": 0.41,
+      "grad_norm": 29.5,
+      "grad_norm_var": 6.380208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2839,
+      "loss/crossentropy": 2.061359938979149,
+      "loss/hidden": 3.3390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18306618463248014,
+      "step": 16400
+    },
+    {
+      "epoch": 0.41025,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.081184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1644,
+      "loss/crossentropy": 2.210022081434727,
+      "loss/hidden": 3.35859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20350072477012873,
+      "step": 16410
+    },
+    {
+      "epoch": 0.4105,
+      "grad_norm": 30.125,
+      "grad_norm_var": 6.44140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3499,
+      "loss/crossentropy": 2.2339374631643296,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19218460842967033,
+      "step": 16420
+    },
+    {
+      "epoch": 0.41075,
+      "grad_norm": 34.5,
+      "grad_norm_var": 7.73515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3014,
+      "loss/crossentropy": 2.259621387720108,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19978805705904962,
+      "step": 16430
+    },
+    {
+      "epoch": 0.411,
+      "grad_norm": 30.875,
+      "grad_norm_var": 4.659309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.29,
+      "loss/crossentropy": 1.9416244141757488,
+      "loss/hidden": 3.282421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16338116079568862,
+      "step": 16440
+    },
+    {
+      "epoch": 0.41125,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.3218098958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2935,
+      "loss/crossentropy": 2.0271296739578246,
+      "loss/hidden": 3.449609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19065151941031216,
+      "step": 16450
+    },
+    {
+      "epoch": 0.4115,
+      "grad_norm": 30.25,
+      "grad_norm_var": 38.69973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2814,
+      "loss/crossentropy": 1.9666823461651801,
+      "loss/hidden": 3.366015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17561122979968785,
+      "step": 16460
+    },
+    {
+      "epoch": 0.41175,
+      "grad_norm": 39.75,
+      "grad_norm_var": 39.458333333333336,
+      "learning_rate": 0.0001,
+      "loss": 7.3983,
+      "loss/crossentropy": 2.104595158994198,
+      "loss/hidden": 3.347265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18662185575813056,
+      "step": 16470
+    },
+    {
+      "epoch": 0.412,
+      "grad_norm": 29.25,
+      "grad_norm_var": 7.270768229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2818,
+      "loss/crossentropy": 2.118724799156189,
+      "loss/hidden": 3.308984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18437495529651643,
+      "step": 16480
+    },
+    {
+      "epoch": 0.41225,
+      "grad_norm": 29.0,
+      "grad_norm_var": 1.1733723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3057,
+      "loss/crossentropy": 1.988658072054386,
+      "loss/hidden": 3.354296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.188140376098454,
+      "step": 16490
+    },
+    {
+      "epoch": 0.4125,
+      "grad_norm": 45.0,
+      "grad_norm_var": 3.7225065361444265e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3181,
+      "loss/crossentropy": 1.9564834155142308,
+      "loss/hidden": 3.43828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1983109924942255,
+      "step": 16500
+    },
+    {
+      "epoch": 0.41275,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.7225065349305247e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4573,
+      "loss/crossentropy": 2.13476774841547,
+      "loss/hidden": 3.314453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1941231820732355,
+      "step": 16510
+    },
+    {
+      "epoch": 0.413,
+      "grad_norm": 34.5,
+      "grad_norm_var": 20.303059895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.3816,
+      "loss/crossentropy": 1.9623971730470657,
+      "loss/hidden": 3.3828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.186094144359231,
+      "step": 16520
+    },
+    {
+      "epoch": 0.41325,
+      "grad_norm": 30.25,
+      "grad_norm_var": 6.732747395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3583,
+      "loss/crossentropy": 2.065499635785818,
+      "loss/hidden": 3.301171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17725836476311088,
+      "step": 16530
+    },
+    {
+      "epoch": 0.4135,
+      "grad_norm": 31.75,
+      "grad_norm_var": 19.949739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3656,
+      "loss/crossentropy": 2.1521792262792587,
+      "loss/hidden": 3.496875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2034964457154274,
+      "step": 16540
+    },
+    {
+      "epoch": 0.41375,
+      "grad_norm": 29.0,
+      "grad_norm_var": 34.993489583333336,
+      "learning_rate": 0.0001,
+      "loss": 7.2803,
+      "loss/crossentropy": 1.9985820479691028,
+      "loss/hidden": 3.36640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1789230139926076,
+      "step": 16550
+    },
+    {
+      "epoch": 0.414,
+      "grad_norm": 30.75,
+      "grad_norm_var": 17.627018229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2555,
+      "loss/crossentropy": 2.094392439723015,
+      "loss/hidden": 3.303125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18661210238933562,
+      "step": 16560
+    },
+    {
+      "epoch": 0.41425,
+      "grad_norm": 29.75,
+      "grad_norm_var": 12.305208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1799,
+      "loss/crossentropy": 2.051650533825159,
+      "loss/hidden": 3.235546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1674873584881425,
+      "step": 16570
+    },
+    {
+      "epoch": 0.4145,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.2372395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.271,
+      "loss/crossentropy": 2.0924292795360087,
+      "loss/hidden": 3.388671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18475022464990615,
+      "step": 16580
+    },
+    {
+      "epoch": 0.41475,
+      "grad_norm": 29.0,
+      "grad_norm_var": 10.5650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2485,
+      "loss/crossentropy": 2.0977213352918627,
+      "loss/hidden": 3.4390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19010566975921392,
+      "step": 16590
+    },
+    {
+      "epoch": 0.415,
+      "grad_norm": 34.0,
+      "grad_norm_var": 20.742122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4038,
+      "loss/crossentropy": 2.0668084505945443,
+      "loss/hidden": 3.35078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1815639895386994,
+      "step": 16600
+    },
+    {
+      "epoch": 0.41525,
+      "grad_norm": 28.875,
+      "grad_norm_var": 10.555208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1948,
+      "loss/crossentropy": 1.973907208442688,
+      "loss/hidden": 3.30078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17607561368495225,
+      "step": 16610
+    },
+    {
+      "epoch": 0.4155,
+      "grad_norm": 39.75,
+      "grad_norm_var": 12.46640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2484,
+      "loss/crossentropy": 2.0116613686084746,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17850100696086885,
+      "step": 16620
+    },
+    {
+      "epoch": 0.41575,
+      "grad_norm": 28.5,
+      "grad_norm_var": 13.439322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2806,
+      "loss/crossentropy": 2.14372631162405,
+      "loss/hidden": 3.3890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1865269223228097,
+      "step": 16630
+    },
+    {
+      "epoch": 0.416,
+      "grad_norm": 30.875,
+      "grad_norm_var": 41.07604166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3744,
+      "loss/crossentropy": 2.0837054416537284,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17686022398993373,
+      "step": 16640
+    },
+    {
+      "epoch": 0.41625,
+      "grad_norm": 33.0,
+      "grad_norm_var": 4.229622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2909,
+      "loss/crossentropy": 1.874044494330883,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1757553458213806,
+      "step": 16650
+    },
+    {
+      "epoch": 0.4165,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.6275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.1618,
+      "loss/crossentropy": 2.0809514716267588,
+      "loss/hidden": 3.323828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18690248318016528,
+      "step": 16660
+    },
+    {
+      "epoch": 0.41675,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.6587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5271,
+      "loss/crossentropy": 2.1389145731925963,
+      "loss/hidden": 3.408203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19053335022181273,
+      "step": 16670
+    },
+    {
+      "epoch": 0.417,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.4676432291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2686,
+      "loss/crossentropy": 1.9298078820109368,
+      "loss/hidden": 3.4265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1739122910425067,
+      "step": 16680
+    },
+    {
+      "epoch": 0.41725,
+      "grad_norm": 29.0,
+      "grad_norm_var": 1.7457682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2261,
+      "loss/crossentropy": 1.8423747673630715,
+      "loss/hidden": 3.376953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1881049584597349,
+      "step": 16690
+    },
+    {
+      "epoch": 0.4175,
+      "grad_norm": 29.25,
+      "grad_norm_var": 2.6733723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3344,
+      "loss/crossentropy": 2.116480963677168,
+      "loss/hidden": 3.32734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20197081826627256,
+      "step": 16700
+    },
+    {
+      "epoch": 0.41775,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.7705729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3253,
+      "loss/crossentropy": 2.0403254240751267,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17736360374838114,
+      "step": 16710
+    },
+    {
+      "epoch": 0.418,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.7622395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.37,
+      "loss/crossentropy": 2.1124753206968307,
+      "loss/hidden": 3.284765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17344876211136578,
+      "step": 16720
+    },
+    {
+      "epoch": 0.41825,
+      "grad_norm": 30.25,
+      "grad_norm_var": 4.176497395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4396,
+      "loss/crossentropy": 2.163897532224655,
+      "loss/hidden": 3.484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20297291725873948,
+      "step": 16730
+    },
+    {
+      "epoch": 0.4185,
+      "grad_norm": 27.75,
+      "grad_norm_var": 6.505989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1872,
+      "loss/crossentropy": 2.2656417414546013,
+      "loss/hidden": 3.297265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18971311207860708,
+      "step": 16740
+    },
+    {
+      "epoch": 0.41875,
+      "grad_norm": 29.25,
+      "grad_norm_var": 5.057291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3375,
+      "loss/crossentropy": 2.0005149811506273,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1829688923433423,
+      "step": 16750
+    },
+    {
+      "epoch": 0.419,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.914322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2354,
+      "loss/crossentropy": 2.1137810915708544,
+      "loss/hidden": 3.371875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19180109687149524,
+      "step": 16760
+    },
+    {
+      "epoch": 0.41925,
+      "grad_norm": 34.25,
+      "grad_norm_var": 3.4905598958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2489,
+      "loss/crossentropy": 2.1259815394878387,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18057906962931156,
+      "step": 16770
+    },
+    {
+      "epoch": 0.4195,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.396875,
+      "learning_rate": 0.0001,
+      "loss": 7.2524,
+      "loss/crossentropy": 1.9959501296281814,
+      "loss/hidden": 3.390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18154409490525722,
+      "step": 16780
+    },
+    {
+      "epoch": 0.41975,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.814322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1953,
+      "loss/crossentropy": 2.0803288385272025,
+      "loss/hidden": 3.296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17842694241553544,
+      "step": 16790
+    },
+    {
+      "epoch": 0.42,
+      "grad_norm": 34.0,
+      "grad_norm_var": 4.6087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3088,
+      "loss/crossentropy": 2.1200789496302606,
+      "loss/hidden": 3.381640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18628805093467235,
+      "step": 16800
+    },
+    {
+      "epoch": 0.42025,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.8666666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3824,
+      "loss/crossentropy": 2.1946793287992477,
+      "loss/hidden": 3.415234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18889498952776193,
+      "step": 16810
+    },
+    {
+      "epoch": 0.4205,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.343489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2866,
+      "loss/crossentropy": 2.147686219215393,
+      "loss/hidden": 3.257421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17540164943784475,
+      "step": 16820
+    },
+    {
+      "epoch": 0.42075,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.631705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3552,
+      "loss/crossentropy": 2.2283954128623007,
+      "loss/hidden": 3.301953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1934325246140361,
+      "step": 16830
+    },
+    {
+      "epoch": 0.421,
+      "grad_norm": 28.75,
+      "grad_norm_var": 1.9114583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.285,
+      "loss/crossentropy": 2.0364210322499274,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18664334248751402,
+      "step": 16840
+    },
+    {
+      "epoch": 0.42125,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.8322916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3212,
+      "loss/crossentropy": 2.166780227422714,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18862007781863213,
+      "step": 16850
+    },
+    {
+      "epoch": 0.4215,
+      "grad_norm": 28.375,
+      "grad_norm_var": 1.6910807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2594,
+      "loss/crossentropy": 1.9404787957668304,
+      "loss/hidden": 3.332421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1732590550556779,
+      "step": 16860
+    },
+    {
+      "epoch": 0.42175,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.4895182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2446,
+      "loss/crossentropy": 2.146845671534538,
+      "loss/hidden": 3.340625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19646301120519638,
+      "step": 16870
+    },
+    {
+      "epoch": 0.422,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.0452473958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3155,
+      "loss/crossentropy": 2.0759237363934515,
+      "loss/hidden": 3.414453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2006880532950163,
+      "step": 16880
+    },
+    {
+      "epoch": 0.42225,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.7358723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2928,
+      "loss/crossentropy": 2.1153281182050705,
+      "loss/hidden": 3.35546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1919462438672781,
+      "step": 16890
+    },
+    {
+      "epoch": 0.4225,
+      "grad_norm": 34.0,
+      "grad_norm_var": 3.1927083333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3425,
+      "loss/crossentropy": 2.128311688452959,
+      "loss/hidden": 3.31328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1898970402777195,
+      "step": 16900
+    },
+    {
+      "epoch": 0.42275,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.2884765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3652,
+      "loss/crossentropy": 2.14291021078825,
+      "loss/hidden": 3.300390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1799526395276189,
+      "step": 16910
+    },
+    {
+      "epoch": 0.423,
+      "grad_norm": 35.75,
+      "grad_norm_var": 2.9853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.374,
+      "loss/crossentropy": 2.172788438200951,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19616848770529033,
+      "step": 16920
+    },
+    {
+      "epoch": 0.42325,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.1434895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.256,
+      "loss/crossentropy": 1.9609168127179146,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19055727850645782,
+      "step": 16930
+    },
+    {
+      "epoch": 0.4235,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.7197265625,
+      "learning_rate": 0.0001,
+      "loss": 7.337,
+      "loss/crossentropy": 2.3514596074819565,
+      "loss/hidden": 3.26875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18605006970465182,
+      "step": 16940
+    },
+    {
+      "epoch": 0.42375,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.0171223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3496,
+      "loss/crossentropy": 2.125219741463661,
+      "loss/hidden": 3.36015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19694697633385658,
+      "step": 16950
+    },
+    {
+      "epoch": 0.424,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.827018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2494,
+      "loss/crossentropy": 2.1004397854208947,
+      "loss/hidden": 3.37421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19005921352654695,
+      "step": 16960
+    },
+    {
+      "epoch": 0.42425,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.16875,
+      "learning_rate": 0.0001,
+      "loss": 7.3353,
+      "loss/crossentropy": 2.2350980192422867,
+      "loss/hidden": 3.389453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20827819555997848,
+      "step": 16970
+    },
+    {
+      "epoch": 0.4245,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.7139973958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2948,
+      "loss/crossentropy": 2.057795172929764,
+      "loss/hidden": 3.41015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18584750667214395,
+      "step": 16980
+    },
+    {
+      "epoch": 0.42475,
+      "grad_norm": 33.5,
+      "grad_norm_var": 9.813541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2737,
+      "loss/crossentropy": 1.9335192561149597,
+      "loss/hidden": 3.38984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17950817625969648,
+      "step": 16990
+    },
+    {
+      "epoch": 0.425,
+      "grad_norm": 30.625,
+      "grad_norm_var": 8.45,
+      "learning_rate": 0.0001,
+      "loss": 7.2137,
+      "loss/crossentropy": 2.158860814571381,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19252809565514326,
+      "step": 17000
+    },
+    {
+      "epoch": 0.42525,
+      "grad_norm": 34.75,
+      "grad_norm_var": 3.940559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3198,
+      "loss/crossentropy": 1.9835775896906853,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18570326836779713,
+      "step": 17010
+    },
+    {
+      "epoch": 0.4255,
+      "grad_norm": 30.625,
+      "grad_norm_var": 13.29140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4143,
+      "loss/crossentropy": 1.9494660779833795,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2034120644442737,
+      "step": 17020
+    },
+    {
+      "epoch": 0.42575,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.7135416666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2964,
+      "loss/crossentropy": 2.1227547653019427,
+      "loss/hidden": 3.33828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18930197739973664,
+      "step": 17030
+    },
+    {
+      "epoch": 0.426,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.07890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2105,
+      "loss/crossentropy": 1.9439098060131073,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1776129573583603,
+      "step": 17040
+    },
+    {
+      "epoch": 0.42625,
+      "grad_norm": 43.25,
+      "grad_norm_var": 25.977018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3636,
+      "loss/crossentropy": 2.1215348944067953,
+      "loss/hidden": 3.366015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19092475529760122,
+      "step": 17050
+    },
+    {
+      "epoch": 0.4265,
+      "grad_norm": 32.75,
+      "grad_norm_var": 15.020572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3927,
+      "loss/crossentropy": 2.0381774201989176,
+      "loss/hidden": 3.33046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18921099193394184,
+      "step": 17060
+    },
+    {
+      "epoch": 0.42675,
+      "grad_norm": 34.0,
+      "grad_norm_var": 3.159309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2753,
+      "loss/crossentropy": 2.053529141843319,
+      "loss/hidden": 3.2734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18555910978466272,
+      "step": 17070
+    },
+    {
+      "epoch": 0.427,
+      "grad_norm": 28.875,
+      "grad_norm_var": 10.44140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3331,
+      "loss/crossentropy": 2.0321754805743693,
+      "loss/hidden": 3.325,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17106994595378638,
+      "step": 17080
+    },
+    {
+      "epoch": 0.42725,
+      "grad_norm": 35.75,
+      "grad_norm_var": 11.438541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2905,
+      "loss/crossentropy": 1.9504701748490334,
+      "loss/hidden": 3.262890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17157490737736225,
+      "step": 17090
+    },
+    {
+      "epoch": 0.4275,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.651497395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4396,
+      "loss/crossentropy": 2.1505665600299837,
+      "loss/hidden": 3.490625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20605666618794202,
+      "step": 17100
+    },
+    {
+      "epoch": 0.42775,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.6688651155212861e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3328,
+      "loss/crossentropy": 2.0673328816890715,
+      "loss/hidden": 3.362109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17912605479359628,
+      "step": 17110
+    },
+    {
+      "epoch": 0.428,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.7436848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3111,
+      "loss/crossentropy": 2.0934616670012476,
+      "loss/hidden": 3.43203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19241752084344627,
+      "step": 17120
+    },
+    {
+      "epoch": 0.42825,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.4580729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2911,
+      "loss/crossentropy": 2.0029770001769065,
+      "loss/hidden": 3.325390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1792398665100336,
+      "step": 17130
+    },
+    {
+      "epoch": 0.4285,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.4184895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3235,
+      "loss/crossentropy": 2.232249027490616,
+      "loss/hidden": 3.354296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1950904905796051,
+      "step": 17140
+    },
+    {
+      "epoch": 0.42875,
+      "grad_norm": 31.0,
+      "grad_norm_var": 5.405143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3339,
+      "loss/crossentropy": 2.1042421609163284,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18525240514427424,
+      "step": 17150
+    },
+    {
+      "epoch": 0.429,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.44765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3841,
+      "loss/crossentropy": 2.0857191674411295,
+      "loss/hidden": 3.331640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18139904439449311,
+      "step": 17160
+    },
+    {
+      "epoch": 0.42925,
+      "grad_norm": 26.75,
+      "grad_norm_var": 9.897330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2916,
+      "loss/crossentropy": 1.9972345970571042,
+      "loss/hidden": 3.359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19361406620591878,
+      "step": 17170
+    },
+    {
+      "epoch": 0.4295,
+      "grad_norm": 33.0,
+      "grad_norm_var": 11.613997395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3454,
+      "loss/crossentropy": 2.0664907500147818,
+      "loss/hidden": 3.33203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18069064132869245,
+      "step": 17180
+    },
+    {
+      "epoch": 0.42975,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.7225065363373645e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3382,
+      "loss/crossentropy": 2.174539531767368,
+      "loss/hidden": 3.417578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1991014001891017,
+      "step": 17190
+    },
+    {
+      "epoch": 0.43,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.0708333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3096,
+      "loss/crossentropy": 1.952410513907671,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1856127019971609,
+      "step": 17200
+    },
+    {
+      "epoch": 0.43025,
+      "grad_norm": 38.75,
+      "grad_norm_var": 8.0275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4419,
+      "loss/crossentropy": 2.128667399287224,
+      "loss/hidden": 3.334375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2061656942591071,
+      "step": 17210
+    },
+    {
+      "epoch": 0.4305,
+      "grad_norm": 30.125,
+      "grad_norm_var": 8.792708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2925,
+      "loss/crossentropy": 1.969744972884655,
+      "loss/hidden": 3.486328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2011998776346445,
+      "step": 17220
+    },
+    {
+      "epoch": 0.43075,
+      "grad_norm": 49.75,
+      "grad_norm_var": 25.044791666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3512,
+      "loss/crossentropy": 2.09512839615345,
+      "loss/hidden": 3.272265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17890227846801282,
+      "step": 17230
+    },
+    {
+      "epoch": 0.431,
+      "grad_norm": 29.625,
+      "grad_norm_var": 25.836458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3174,
+      "loss/crossentropy": 2.2069885596632957,
+      "loss/hidden": 3.38046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20973877012729644,
+      "step": 17240
+    },
+    {
+      "epoch": 0.43125,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.7309405650762245e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4689,
+      "loss/crossentropy": 2.197174680233002,
+      "loss/hidden": 3.248828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.184923891723156,
+      "step": 17250
+    },
+    {
+      "epoch": 0.4315,
+      "grad_norm": 28.0,
+      "grad_norm_var": 2.730940565530678e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2989,
+      "loss/crossentropy": 2.08426301702857,
+      "loss/hidden": 3.3953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18929961118847133,
+      "step": 17260
+    },
+    {
+      "epoch": 0.43175,
+      "grad_norm": 31.75,
+      "grad_norm_var": 9.408268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3047,
+      "loss/crossentropy": 1.9928709417581558,
+      "loss/hidden": 3.44375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19445511922240258,
+      "step": 17270
+    },
+    {
+      "epoch": 0.432,
+      "grad_norm": 41.25,
+      "grad_norm_var": 17.242708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2872,
+      "loss/crossentropy": 1.9035924103111028,
+      "loss/hidden": 3.391015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18102165032178164,
+      "step": 17280
+    },
+    {
+      "epoch": 0.43225,
+      "grad_norm": 28.125,
+      "grad_norm_var": 14.945768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3079,
+      "loss/crossentropy": 2.2340116620063784,
+      "loss/hidden": 3.29921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19132067933678626,
+      "step": 17290
+    },
+    {
+      "epoch": 0.4325,
+      "grad_norm": 29.5,
+      "grad_norm_var": 8.06875,
+      "learning_rate": 0.0001,
+      "loss": 7.2915,
+      "loss/crossentropy": 2.0194656267762183,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19431380424648523,
+      "step": 17300
+    },
+    {
+      "epoch": 0.43275,
+      "grad_norm": 30.875,
+      "grad_norm_var": 8.566666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.261,
+      "loss/crossentropy": 2.001611000299454,
+      "loss/hidden": 3.28828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17275465354323388,
+      "step": 17310
+    },
+    {
+      "epoch": 0.433,
+      "grad_norm": 35.25,
+      "grad_norm_var": 13.3775390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3377,
+      "loss/crossentropy": 2.114110495150089,
+      "loss/hidden": 3.439453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20012194942682981,
+      "step": 17320
+    },
+    {
+      "epoch": 0.43325,
+      "grad_norm": 39.0,
+      "grad_norm_var": 44.28958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3729,
+      "loss/crossentropy": 2.2978450536727903,
+      "loss/hidden": 3.27734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18779626674950123,
+      "step": 17330
+    },
+    {
+      "epoch": 0.4335,
+      "grad_norm": 32.0,
+      "grad_norm_var": 33.8791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3487,
+      "loss/crossentropy": 2.0877898022532464,
+      "loss/hidden": 3.296484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18020253200083972,
+      "step": 17340
+    },
+    {
+      "epoch": 0.43375,
+      "grad_norm": 28.25,
+      "grad_norm_var": 6.508333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2337,
+      "loss/crossentropy": 1.9320446044206618,
+      "loss/hidden": 3.407421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18643499026075006,
+      "step": 17350
+    },
+    {
+      "epoch": 0.434,
+      "grad_norm": 38.25,
+      "grad_norm_var": 9.514322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3395,
+      "loss/crossentropy": 2.033699007332325,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18971308488398791,
+      "step": 17360
+    },
+    {
+      "epoch": 0.43425,
+      "grad_norm": 30.0,
+      "grad_norm_var": 6.888997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.353,
+      "loss/crossentropy": 2.1282121136784555,
+      "loss/hidden": 3.30546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1824582252651453,
+      "step": 17370
+    },
+    {
+      "epoch": 0.4345,
+      "grad_norm": 30.5,
+      "grad_norm_var": 4.592708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2093,
+      "loss/crossentropy": 2.1514922633767126,
+      "loss/hidden": 3.37421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19244836810976268,
+      "step": 17380
+    },
+    {
+      "epoch": 0.43475,
+      "grad_norm": 31.0,
+      "grad_norm_var": 8.587434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.28,
+      "loss/crossentropy": 1.9829683952033519,
+      "loss/hidden": 3.3828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1997856667265296,
+      "step": 17390
+    },
+    {
+      "epoch": 0.435,
+      "grad_norm": 30.25,
+      "grad_norm_var": 12.732747395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.327,
+      "loss/crossentropy": 2.0963875532150267,
+      "loss/hidden": 3.2953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18260720651596785,
+      "step": 17400
+    },
+    {
+      "epoch": 0.43525,
+      "grad_norm": 39.25,
+      "grad_norm_var": 13.623893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2851,
+      "loss/crossentropy": 2.1481306463479997,
+      "loss/hidden": 3.328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18422619421035052,
+      "step": 17410
+    },
+    {
+      "epoch": 0.4355,
+      "grad_norm": 35.25,
+      "grad_norm_var": 33.15514322916667,
+      "learning_rate": 0.0001,
+      "loss": 7.2614,
+      "loss/crossentropy": 1.94967889636755,
+      "loss/hidden": 3.31015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16873782277107238,
+      "step": 17420
+    },
+    {
+      "epoch": 0.43575,
+      "grad_norm": 30.25,
+      "grad_norm_var": 19.037239583333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2782,
+      "loss/crossentropy": 2.3510662406682967,
+      "loss/hidden": 3.228515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18572253845632075,
+      "step": 17430
+    },
+    {
+      "epoch": 0.436,
+      "grad_norm": 28.375,
+      "grad_norm_var": 1.3291666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3387,
+      "loss/crossentropy": 2.1440586492419245,
+      "loss/hidden": 3.355078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1921809207648039,
+      "step": 17440
+    },
+    {
+      "epoch": 0.43625,
+      "grad_norm": 30.0,
+      "grad_norm_var": 1.2837890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2998,
+      "loss/crossentropy": 2.045154483616352,
+      "loss/hidden": 3.39140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20081294253468512,
+      "step": 17450
+    },
+    {
+      "epoch": 0.4365,
+      "grad_norm": 35.75,
+      "grad_norm_var": 4.692643229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2213,
+      "loss/crossentropy": 1.9950920060276984,
+      "loss/hidden": 3.28828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18348362799733878,
+      "step": 17460
+    },
+    {
+      "epoch": 0.43675,
+      "grad_norm": 28.625,
+      "grad_norm_var": 47.77890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2084,
+      "loss/crossentropy": 2.0935964420437814,
+      "loss/hidden": 3.390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1935350162908435,
+      "step": 17470
+    },
+    {
+      "epoch": 0.437,
+      "grad_norm": 30.375,
+      "grad_norm_var": 51.95462239583333,
+      "learning_rate": 0.0001,
+      "loss": 7.3469,
+      "loss/crossentropy": 2.2382303804159163,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2008738409727812,
+      "step": 17480
+    },
+    {
+      "epoch": 0.43725,
+      "grad_norm": 30.75,
+      "grad_norm_var": 7.177083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2088,
+      "loss/crossentropy": 2.0825397919863464,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19918248979374767,
+      "step": 17490
+    },
+    {
+      "epoch": 0.4375,
+      "grad_norm": 31.0,
+      "grad_norm_var": 41.16015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3124,
+      "loss/crossentropy": 2.1086046427488325,
+      "loss/hidden": 3.40703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2059457702562213,
+      "step": 17500
+    },
+    {
+      "epoch": 0.43775,
+      "grad_norm": 30.25,
+      "grad_norm_var": 41.1275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3142,
+      "loss/crossentropy": 2.1939937084913255,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19669633246958257,
+      "step": 17510
+    },
+    {
+      "epoch": 0.438,
+      "grad_norm": 30.0,
+      "grad_norm_var": 5.426822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4971,
+      "loss/crossentropy": 2.077922207117081,
+      "loss/hidden": 3.394140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19905825443565844,
+      "step": 17520
+    },
+    {
+      "epoch": 0.43825,
+      "grad_norm": 35.0,
+      "grad_norm_var": 5.139518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3389,
+      "loss/crossentropy": 2.2604495763778685,
+      "loss/hidden": 3.26796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18292145486921071,
+      "step": 17530
+    },
+    {
+      "epoch": 0.4385,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.9483723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3261,
+      "loss/crossentropy": 2.033947843313217,
+      "loss/hidden": 3.444140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18323329892009496,
+      "step": 17540
+    },
+    {
+      "epoch": 0.43875,
+      "grad_norm": 30.625,
+      "grad_norm_var": 0.8455729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2981,
+      "loss/crossentropy": 2.2359601676464083,
+      "loss/hidden": 3.284765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18464031405746936,
+      "step": 17550
+    },
+    {
+      "epoch": 0.439,
+      "grad_norm": 40.5,
+      "grad_norm_var": 2.3053504042993976e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3321,
+      "loss/crossentropy": 1.9767519772052764,
+      "loss/hidden": 3.375390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17584269326180219,
+      "step": 17560
+    },
+    {
+      "epoch": 0.43925,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.3053504048877535e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2311,
+      "loss/crossentropy": 2.08755609691143,
+      "loss/hidden": 3.34921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18783280439674854,
+      "step": 17570
+    },
+    {
+      "epoch": 0.4395,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.872330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2714,
+      "loss/crossentropy": 2.1911805018782617,
+      "loss/hidden": 3.3390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19903278667479754,
+      "step": 17580
+    },
+    {
+      "epoch": 0.43975,
+      "grad_norm": 28.0,
+      "grad_norm_var": 16.263997395833332,
+      "learning_rate": 0.0001,
+      "loss": 7.4407,
+      "loss/crossentropy": 2.1172087833285334,
+      "loss/hidden": 3.417578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1971781937405467,
+      "step": 17590
+    },
+    {
+      "epoch": 0.44,
+      "grad_norm": 31.25,
+      "grad_norm_var": 5.95390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4514,
+      "loss/crossentropy": 2.0933860525488854,
+      "loss/hidden": 3.54453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21392161287367345,
+      "step": 17600
+    },
+    {
+      "epoch": 0.44025,
+      "grad_norm": 30.375,
+      "grad_norm_var": 8.245833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3118,
+      "loss/crossentropy": 2.1100318849086763,
+      "loss/hidden": 3.432421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21126667950302364,
+      "step": 17610
+    },
+    {
+      "epoch": 0.4405,
+      "grad_norm": 29.0,
+      "grad_norm_var": 7.802083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2395,
+      "loss/crossentropy": 1.9662518702447414,
+      "loss/hidden": 3.31640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17263798629865051,
+      "step": 17620
+    },
+    {
+      "epoch": 0.44075,
+      "grad_norm": 37.75,
+      "grad_norm_var": 21.355208333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3091,
+      "loss/crossentropy": 2.1022576719522474,
+      "loss/hidden": 3.41328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18059359192848207,
+      "step": 17630
+    },
+    {
+      "epoch": 0.441,
+      "grad_norm": 29.0,
+      "grad_norm_var": 18.4228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2647,
+      "loss/crossentropy": 2.158071845769882,
+      "loss/hidden": 3.255078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17066741604357957,
+      "step": 17640
+    },
+    {
+      "epoch": 0.44125,
+      "grad_norm": 36.5,
+      "grad_norm_var": 21.584309895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.3432,
+      "loss/crossentropy": 2.017947067320347,
+      "loss/hidden": 3.37109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1873489862307906,
+      "step": 17650
+    },
+    {
+      "epoch": 0.4415,
+      "grad_norm": 33.0,
+      "grad_norm_var": 6.862239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2728,
+      "loss/crossentropy": 2.2292807698249817,
+      "loss/hidden": 3.309765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18392078932374717,
+      "step": 17660
+    },
+    {
+      "epoch": 0.44175,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.849739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1737,
+      "loss/crossentropy": 2.1030597440898418,
+      "loss/hidden": 3.316796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18242286536842583,
+      "step": 17670
+    },
+    {
+      "epoch": 0.442,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.0385416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3015,
+      "loss/crossentropy": 2.0946084149181843,
+      "loss/hidden": 3.308203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1821262989193201,
+      "step": 17680
+    },
+    {
+      "epoch": 0.44225,
+      "grad_norm": 35.0,
+      "grad_norm_var": 3.60390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2853,
+      "loss/crossentropy": 2.151443210244179,
+      "loss/hidden": 3.39453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19624145627021788,
+      "step": 17690
+    },
+    {
+      "epoch": 0.4425,
+      "grad_norm": 34.5,
+      "grad_norm_var": 3.2958333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4457,
+      "loss/crossentropy": 2.153268966078758,
+      "loss/hidden": 3.287890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1952244239859283,
+      "step": 17700
+    },
+    {
+      "epoch": 0.44275,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.9791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2833,
+      "loss/crossentropy": 2.0192960262298585,
+      "loss/hidden": 3.27578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1719641860574484,
+      "step": 17710
+    },
+    {
+      "epoch": 0.443,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.4875,
+      "learning_rate": 0.0001,
+      "loss": 7.3388,
+      "loss/crossentropy": 2.0337466023862363,
+      "loss/hidden": 3.32890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18310964982956648,
+      "step": 17720
+    },
+    {
+      "epoch": 0.44325,
+      "grad_norm": 27.5,
+      "grad_norm_var": 8.132747395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2481,
+      "loss/crossentropy": 2.083891141414642,
+      "loss/hidden": 3.398046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1840117072686553,
+      "step": 17730
+    },
+    {
+      "epoch": 0.4435,
+      "grad_norm": 65.0,
+      "grad_norm_var": 79.64270833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2557,
+      "loss/crossentropy": 2.193793588876724,
+      "loss/hidden": 3.2828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18301294278353453,
+      "step": 17740
+    },
+    {
+      "epoch": 0.44375,
+      "grad_norm": 31.25,
+      "grad_norm_var": 79.53098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3745,
+      "loss/crossentropy": 2.1639930069446565,
+      "loss/hidden": 3.329296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18582290820777417,
+      "step": 17750
+    },
+    {
+      "epoch": 0.444,
+      "grad_norm": 28.875,
+      "grad_norm_var": 4.594791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4009,
+      "loss/crossentropy": 2.0524846225976945,
+      "loss/hidden": 3.290234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18291038312017918,
+      "step": 17760
+    },
+    {
+      "epoch": 0.44425,
+      "grad_norm": 29.25,
+      "grad_norm_var": 5.265559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2015,
+      "loss/crossentropy": 2.0137507632374763,
+      "loss/hidden": 3.365625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17316031139343976,
+      "step": 17770
+    },
+    {
+      "epoch": 0.4445,
+      "grad_norm": 32.0,
+      "grad_norm_var": 4.96015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3163,
+      "loss/crossentropy": 2.172424706816673,
+      "loss/hidden": 3.40546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19778120778501035,
+      "step": 17780
+    },
+    {
+      "epoch": 0.44475,
+      "grad_norm": 29.25,
+      "grad_norm_var": 217.27083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1559,
+      "loss/crossentropy": 2.076275750249624,
+      "loss/hidden": 3.3234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17568769995123149,
+      "step": 17790
+    },
+    {
+      "epoch": 0.445,
+      "grad_norm": 33.25,
+      "grad_norm_var": 223.44368489583334,
+      "learning_rate": 0.0001,
+      "loss": 7.3438,
+      "loss/crossentropy": 2.270902451872826,
+      "loss/hidden": 3.22265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1801173485815525,
+      "step": 17800
+    },
+    {
+      "epoch": 0.44525,
+      "grad_norm": 39.75,
+      "grad_norm_var": 12.383072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2998,
+      "loss/crossentropy": 2.1936209172010424,
+      "loss/hidden": 3.49296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19792029969394206,
+      "step": 17810
+    },
+    {
+      "epoch": 0.4455,
+      "grad_norm": 29.5,
+      "grad_norm_var": 7.808268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3011,
+      "loss/crossentropy": 2.1165911227464678,
+      "loss/hidden": 3.303515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1850597882643342,
+      "step": 17820
+    },
+    {
+      "epoch": 0.44575,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.218489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2305,
+      "loss/crossentropy": 2.0913813322782517,
+      "loss/hidden": 3.26171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17276749946177006,
+      "step": 17830
+    },
+    {
+      "epoch": 0.446,
+      "grad_norm": 48.25,
+      "grad_norm_var": 40.351497395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3693,
+      "loss/crossentropy": 2.014899070560932,
+      "loss/hidden": 3.358984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2062089815735817,
+      "step": 17840
+    },
+    {
+      "epoch": 0.44625,
+      "grad_norm": 32.5,
+      "grad_norm_var": 31.93515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3685,
+      "loss/crossentropy": 2.048551322519779,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17983095441013575,
+      "step": 17850
+    },
+    {
+      "epoch": 0.4465,
+      "grad_norm": 28.75,
+      "grad_norm_var": 30.507291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3944,
+      "loss/crossentropy": 2.094169969111681,
+      "loss/hidden": 3.36171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18028030041605234,
+      "step": 17860
+    },
+    {
+      "epoch": 0.44675,
+      "grad_norm": 46.0,
+      "grad_norm_var": 42.3962890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2828,
+      "loss/crossentropy": 2.001505134999752,
+      "loss/hidden": 3.404296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19261684585362673,
+      "step": 17870
+    },
+    {
+      "epoch": 0.447,
+      "grad_norm": 29.125,
+      "grad_norm_var": 16.0697265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2928,
+      "loss/crossentropy": 2.034270917624235,
+      "loss/hidden": 3.36484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17677650935947894,
+      "step": 17880
+    },
+    {
+      "epoch": 0.44725,
+      "grad_norm": 30.5,
+      "grad_norm_var": 7.384830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2414,
+      "loss/crossentropy": 2.177957884967327,
+      "loss/hidden": 3.348828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20077546443790198,
+      "step": 17890
+    },
+    {
+      "epoch": 0.4475,
+      "grad_norm": 26.5,
+      "grad_norm_var": 19.184309895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2819,
+      "loss/crossentropy": 2.130458039045334,
+      "loss/hidden": 3.44921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19989439938217402,
+      "step": 17900
+    },
+    {
+      "epoch": 0.44775,
+      "grad_norm": 35.5,
+      "grad_norm_var": 44.08515625,
+      "learning_rate": 0.0001,
+      "loss": 7.1854,
+      "loss/crossentropy": 2.1698092609643935,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18095692303031682,
+      "step": 17910
+    },
+    {
+      "epoch": 0.448,
+      "grad_norm": 38.25,
+      "grad_norm_var": 44.31041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3364,
+      "loss/crossentropy": 2.06129896491766,
+      "loss/hidden": 3.30859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18485308829694985,
+      "step": 17920
+    },
+    {
+      "epoch": 0.44825,
+      "grad_norm": 37.25,
+      "grad_norm_var": 30.994205729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3095,
+      "loss/crossentropy": 1.9626867271959783,
+      "loss/hidden": 3.460546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18522640708833932,
+      "step": 17930
+    },
+    {
+      "epoch": 0.4485,
+      "grad_norm": 38.75,
+      "grad_norm_var": 14.283072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3206,
+      "loss/crossentropy": 2.061312362551689,
+      "loss/hidden": 3.4265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18516826909035444,
+      "step": 17940
+    },
+    {
+      "epoch": 0.44875,
+      "grad_norm": 39.25,
+      "grad_norm_var": 12.702083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2801,
+      "loss/crossentropy": 2.0153495140373705,
+      "loss/hidden": 3.389453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1914671964943409,
+      "step": 17950
+    },
+    {
+      "epoch": 0.449,
+      "grad_norm": 30.75,
+      "grad_norm_var": 10.4650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3371,
+      "loss/crossentropy": 2.0574869602918624,
+      "loss/hidden": 3.451953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20203660018742084,
+      "step": 17960
+    },
+    {
+      "epoch": 0.44925,
+      "grad_norm": 31.0,
+      "grad_norm_var": 7.4431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.301,
+      "loss/crossentropy": 2.1529300838708876,
+      "loss/hidden": 3.304296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18314522188156843,
+      "step": 17970
+    },
+    {
+      "epoch": 0.4495,
+      "grad_norm": 29.375,
+      "grad_norm_var": 28.781705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.286,
+      "loss/crossentropy": 2.2260805390775205,
+      "loss/hidden": 3.29296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18513377662748098,
+      "step": 17980
+    },
+    {
+      "epoch": 0.44975,
+      "grad_norm": 29.5,
+      "grad_norm_var": 16.774739583333332,
+      "learning_rate": 0.0001,
+      "loss": 7.4066,
+      "loss/crossentropy": 1.9546913027763366,
+      "loss/hidden": 3.3421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18548114076256753,
+      "step": 17990
+    },
+    {
+      "epoch": 0.45,
+      "grad_norm": 31.875,
+      "grad_norm_var": 14.9619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2574,
+      "loss/crossentropy": 2.0575848095119,
+      "loss/hidden": 3.268359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17614138405770063,
+      "step": 18000
+    },
+    {
+      "epoch": 0.45025,
+      "grad_norm": 36.5,
+      "grad_norm_var": 9.5603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3801,
+      "loss/crossentropy": 2.0923216179013253,
+      "loss/hidden": 3.265234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.180926413834095,
+      "step": 18010
+    },
+    {
+      "epoch": 0.4505,
+      "grad_norm": 27.25,
+      "grad_norm_var": 13.5884765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3101,
+      "loss/crossentropy": 2.0365143597126005,
+      "loss/hidden": 3.3375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1913832610473037,
+      "step": 18020
+    },
+    {
+      "epoch": 0.45075,
+      "grad_norm": 31.875,
+      "grad_norm_var": 16.5962890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2971,
+      "loss/crossentropy": 2.0645697891712187,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19654379645362496,
+      "step": 18030
+    },
+    {
+      "epoch": 0.451,
+      "grad_norm": 29.5,
+      "grad_norm_var": 11.794205729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2328,
+      "loss/crossentropy": 2.0605416260659695,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18139277547597885,
+      "step": 18040
+    },
+    {
+      "epoch": 0.45125,
+      "grad_norm": 32.0,
+      "grad_norm_var": 75.465625,
+      "learning_rate": 0.0001,
+      "loss": 7.2392,
+      "loss/crossentropy": 2.083356872946024,
+      "loss/hidden": 3.34296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2048046118579805,
+      "step": 18050
+    },
+    {
+      "epoch": 0.4515,
+      "grad_norm": 36.5,
+      "grad_norm_var": 35.063541666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2963,
+      "loss/crossentropy": 1.9495014771819115,
+      "loss/hidden": 3.3796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17369234804064035,
+      "step": 18060
+    },
+    {
+      "epoch": 0.45175,
+      "grad_norm": 37.5,
+      "grad_norm_var": 10.093684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2651,
+      "loss/crossentropy": 2.124495878815651,
+      "loss/hidden": 3.360546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19412798956036567,
+      "step": 18070
+    },
+    {
+      "epoch": 0.452,
+      "grad_norm": 26.5,
+      "grad_norm_var": 18.055208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4276,
+      "loss/crossentropy": 2.1983793139457704,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19716158043593168,
+      "step": 18080
+    },
+    {
+      "epoch": 0.45225,
+      "grad_norm": 29.0,
+      "grad_norm_var": 16.8509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2958,
+      "loss/crossentropy": 2.0635679766535757,
+      "loss/hidden": 3.343359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19352352600544692,
+      "step": 18090
+    },
+    {
+      "epoch": 0.4525,
+      "grad_norm": 31.875,
+      "grad_norm_var": 4.707291666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3162,
+      "loss/crossentropy": 2.1419614136219023,
+      "loss/hidden": 3.3515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1909960376098752,
+      "step": 18100
+    },
+    {
+      "epoch": 0.45275,
+      "grad_norm": 28.0,
+      "grad_norm_var": 3.0768229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2917,
+      "loss/crossentropy": 2.0669834718108175,
+      "loss/hidden": 3.356640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18134585060179234,
+      "step": 18110
+    },
+    {
+      "epoch": 0.453,
+      "grad_norm": 29.0,
+      "grad_norm_var": 4.955989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3518,
+      "loss/crossentropy": 2.2067804619669915,
+      "loss/hidden": 3.304296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18476400952786207,
+      "step": 18120
+    },
+    {
+      "epoch": 0.45325,
+      "grad_norm": 30.625,
+      "grad_norm_var": 11.580989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2172,
+      "loss/crossentropy": 2.027407840639353,
+      "loss/hidden": 3.355859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17775150025263428,
+      "step": 18130
+    },
+    {
+      "epoch": 0.4535,
+      "grad_norm": 34.5,
+      "grad_norm_var": 12.3259765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2658,
+      "loss/crossentropy": 2.2659630328416824,
+      "loss/hidden": 3.33046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1918324466794729,
+      "step": 18140
+    },
+    {
+      "epoch": 0.45375,
+      "grad_norm": 28.75,
+      "grad_norm_var": 5.1869140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2627,
+      "loss/crossentropy": 2.1039141535758974,
+      "loss/hidden": 3.3953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18257359908893706,
+      "step": 18150
+    },
+    {
+      "epoch": 0.454,
+      "grad_norm": 28.125,
+      "grad_norm_var": 1.4462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2659,
+      "loss/crossentropy": 2.1540609017014503,
+      "loss/hidden": 3.35625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19397172834724188,
+      "step": 18160
+    },
+    {
+      "epoch": 0.45425,
+      "grad_norm": 28.625,
+      "grad_norm_var": 3.134375,
+      "learning_rate": 0.0001,
+      "loss": 7.2332,
+      "loss/crossentropy": 2.0309683278203012,
+      "loss/hidden": 3.455859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19702901802957057,
+      "step": 18170
+    },
+    {
+      "epoch": 0.4545,
+      "grad_norm": 27.875,
+      "grad_norm_var": 3.4368798057619784e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3245,
+      "loss/crossentropy": 1.9289076425135137,
+      "loss/hidden": 3.287890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.162909129075706,
+      "step": 18180
+    },
+    {
+      "epoch": 0.45475,
+      "grad_norm": 29.75,
+      "grad_norm_var": 6.216666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2699,
+      "loss/crossentropy": 2.1170002311468124,
+      "loss/hidden": 3.242578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17830445962026714,
+      "step": 18190
+    },
+    {
+      "epoch": 0.455,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.5801432291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3454,
+      "loss/crossentropy": 2.1520806990563868,
+      "loss/hidden": 3.418359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.193353005964309,
+      "step": 18200
+    },
+    {
+      "epoch": 0.45525,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.0270182291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2713,
+      "loss/crossentropy": 1.9926446735858918,
+      "loss/hidden": 3.375390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19921180326491594,
+      "step": 18210
+    },
+    {
+      "epoch": 0.4555,
+      "grad_norm": 29.125,
+      "grad_norm_var": 22.040559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3593,
+      "loss/crossentropy": 2.2070549339056016,
+      "loss/hidden": 3.37890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19235619865357875,
+      "step": 18220
+    },
+    {
+      "epoch": 0.45575,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.6927083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3017,
+      "loss/crossentropy": 2.088302744925022,
+      "loss/hidden": 3.290234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17863646000623704,
+      "step": 18230
+    },
+    {
+      "epoch": 0.456,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.7895182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1981,
+      "loss/crossentropy": 2.076305481791496,
+      "loss/hidden": 3.296484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16893827533349395,
+      "step": 18240
+    },
+    {
+      "epoch": 0.45625,
+      "grad_norm": 31.125,
+      "grad_norm_var": 42.75390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3307,
+      "loss/crossentropy": 1.8771976202726364,
+      "loss/hidden": 3.473046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18784729558974506,
+      "step": 18250
+    },
+    {
+      "epoch": 0.4565,
+      "grad_norm": 30.5,
+      "grad_norm_var": 40.77265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2858,
+      "loss/crossentropy": 2.078951980918646,
+      "loss/hidden": 3.365625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19046283271163703,
+      "step": 18260
+    },
+    {
+      "epoch": 0.45675,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.2643229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.296,
+      "loss/crossentropy": 2.1873993456363676,
+      "loss/hidden": 3.309375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18323179706931114,
+      "step": 18270
+    },
+    {
+      "epoch": 0.457,
+      "grad_norm": 27.125,
+      "grad_norm_var": 2.7113932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2805,
+      "loss/crossentropy": 2.1417489528656004,
+      "loss/hidden": 3.2421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1756190422922373,
+      "step": 18280
+    },
+    {
+      "epoch": 0.45725,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.3395182291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3185,
+      "loss/crossentropy": 2.0876102939248087,
+      "loss/hidden": 3.3265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1938194077461958,
+      "step": 18290
+    },
+    {
+      "epoch": 0.4575,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.5551432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3099,
+      "loss/crossentropy": 2.088182657957077,
+      "loss/hidden": 3.28984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18158297892659903,
+      "step": 18300
+    },
+    {
+      "epoch": 0.45775,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.09765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3893,
+      "loss/crossentropy": 2.0980015411973,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19810429476201535,
+      "step": 18310
+    },
+    {
+      "epoch": 0.458,
+      "grad_norm": 32.5,
+      "grad_norm_var": 14.725455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3944,
+      "loss/crossentropy": 2.164955261349678,
+      "loss/hidden": 3.330078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19350772826001048,
+      "step": 18320
+    },
+    {
+      "epoch": 0.45825,
+      "grad_norm": 29.75,
+      "grad_norm_var": 14.030989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3787,
+      "loss/crossentropy": 2.040371149778366,
+      "loss/hidden": 3.290625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17534797079861164,
+      "step": 18330
+    },
+    {
+      "epoch": 0.4585,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.597916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2343,
+      "loss/crossentropy": 1.948985606431961,
+      "loss/hidden": 3.362109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17247573314234615,
+      "step": 18340
+    },
+    {
+      "epoch": 0.45875,
+      "grad_norm": 33.25,
+      "grad_norm_var": 4.514322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2451,
+      "loss/crossentropy": 2.1069004431366922,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2032563455402851,
+      "step": 18350
+    },
+    {
+      "epoch": 0.459,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.6384765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3308,
+      "loss/crossentropy": 2.174188384413719,
+      "loss/hidden": 3.3421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19028009474277496,
+      "step": 18360
+    },
+    {
+      "epoch": 0.45925,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.6384765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3875,
+      "loss/crossentropy": 2.07789751291275,
+      "loss/hidden": 3.314453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1836570143699646,
+      "step": 18370
+    },
+    {
+      "epoch": 0.4595,
+      "grad_norm": 30.5,
+      "grad_norm_var": 9.654622395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2052,
+      "loss/crossentropy": 2.0727937802672387,
+      "loss/hidden": 3.34609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18003296088427306,
+      "step": 18380
+    },
+    {
+      "epoch": 0.45975,
+      "grad_norm": 30.75,
+      "grad_norm_var": 10.531184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.23,
+      "loss/crossentropy": 2.0466937959194182,
+      "loss/hidden": 3.30390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17429272253066302,
+      "step": 18390
+    },
+    {
+      "epoch": 0.46,
+      "grad_norm": 28.5,
+      "grad_norm_var": 5.357291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2161,
+      "loss/crossentropy": 2.007898286730051,
+      "loss/hidden": 3.2984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17548899427056314,
+      "step": 18400
+    },
+    {
+      "epoch": 0.46025,
+      "grad_norm": 30.5,
+      "grad_norm_var": 4.42890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3552,
+      "loss/crossentropy": 2.12602576315403,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18963672500103712,
+      "step": 18410
+    },
+    {
+      "epoch": 0.4605,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.1228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2265,
+      "loss/crossentropy": 1.946645411849022,
+      "loss/hidden": 3.30234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1663317572325468,
+      "step": 18420
+    },
+    {
+      "epoch": 0.46075,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.312239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3208,
+      "loss/crossentropy": 2.0005823224782944,
+      "loss/hidden": 3.2734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1826445495709777,
+      "step": 18430
+    },
+    {
+      "epoch": 0.461,
+      "grad_norm": 30.375,
+      "grad_norm_var": 0.9931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3267,
+      "loss/crossentropy": 2.1357066452503206,
+      "loss/hidden": 3.40859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19712673518806695,
+      "step": 18440
+    },
+    {
+      "epoch": 0.46125,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.9497395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3522,
+      "loss/crossentropy": 2.2581495225429533,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20862563773989679,
+      "step": 18450
+    },
+    {
+      "epoch": 0.4615,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.3629557291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2707,
+      "loss/crossentropy": 1.8819428831338882,
+      "loss/hidden": 3.32890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17904695197939874,
+      "step": 18460
+    },
+    {
+      "epoch": 0.46175,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.0885416666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.441,
+      "loss/crossentropy": 2.1469796687364577,
+      "loss/hidden": 3.4140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20115426667034625,
+      "step": 18470
+    },
+    {
+      "epoch": 0.462,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.2333333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.337,
+      "loss/crossentropy": 2.054457852616906,
+      "loss/hidden": 3.371875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18193845450878143,
+      "step": 18480
+    },
+    {
+      "epoch": 0.46225,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.3489583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2604,
+      "loss/crossentropy": 2.0947310835123063,
+      "loss/hidden": 3.3,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18451766278594733,
+      "step": 18490
+    },
+    {
+      "epoch": 0.4625,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.9104166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3593,
+      "loss/crossentropy": 2.1241430282592773,
+      "loss/hidden": 3.39296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19629997350275516,
+      "step": 18500
+    },
+    {
+      "epoch": 0.46275,
+      "grad_norm": 27.75,
+      "grad_norm_var": 11.044791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3238,
+      "loss/crossentropy": 1.99822306483984,
+      "loss/hidden": 3.34375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17915242053568364,
+      "step": 18510
+    },
+    {
+      "epoch": 0.463,
+      "grad_norm": 28.25,
+      "grad_norm_var": 21.5275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3394,
+      "loss/crossentropy": 1.996624768525362,
+      "loss/hidden": 3.2890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1680696103721857,
+      "step": 18520
+    },
+    {
+      "epoch": 0.46325,
+      "grad_norm": 31.875,
+      "grad_norm_var": 12.8447265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2899,
+      "loss/crossentropy": 1.9818643540143968,
+      "loss/hidden": 3.38046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1780361386016011,
+      "step": 18530
+    },
+    {
+      "epoch": 0.4635,
+      "grad_norm": 28.125,
+      "grad_norm_var": 2.0306640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3093,
+      "loss/crossentropy": 1.976102039217949,
+      "loss/hidden": 3.4140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17688030060380697,
+      "step": 18540
+    },
+    {
+      "epoch": 0.46375,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.6103515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4761,
+      "loss/crossentropy": 2.0188881896436213,
+      "loss/hidden": 3.430078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19428470879793167,
+      "step": 18550
+    },
+    {
+      "epoch": 0.464,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.060416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2272,
+      "loss/crossentropy": 1.997609743475914,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.179104876331985,
+      "step": 18560
+    },
+    {
+      "epoch": 0.46425,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.9645182291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4043,
+      "loss/crossentropy": 2.1501515746116637,
+      "loss/hidden": 3.4203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20029870867729188,
+      "step": 18570
+    },
+    {
+      "epoch": 0.4645,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.39140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3202,
+      "loss/crossentropy": 2.1886572629213332,
+      "loss/hidden": 3.33046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18652867767959833,
+      "step": 18580
+    },
+    {
+      "epoch": 0.46475,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.2582682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.372,
+      "loss/crossentropy": 2.1719513699412345,
+      "loss/hidden": 3.3625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1916284618899226,
+      "step": 18590
+    },
+    {
+      "epoch": 0.465,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.7514973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2974,
+      "loss/crossentropy": 2.013065665960312,
+      "loss/hidden": 3.364453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1881342137232423,
+      "step": 18600
+    },
+    {
+      "epoch": 0.46525,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.730208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3544,
+      "loss/crossentropy": 2.241141265630722,
+      "loss/hidden": 3.438671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19869245793670415,
+      "step": 18610
+    },
+    {
+      "epoch": 0.4655,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.289322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2523,
+      "loss/crossentropy": 2.1553829818964005,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19479922130703925,
+      "step": 18620
+    },
+    {
+      "epoch": 0.46575,
+      "grad_norm": 32.75,
+      "grad_norm_var": 1.896875,
+      "learning_rate": 0.0001,
+      "loss": 7.4636,
+      "loss/crossentropy": 2.1676395773887633,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19261010140180587,
+      "step": 18630
+    },
+    {
+      "epoch": 0.466,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.2739583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3126,
+      "loss/crossentropy": 1.9635831721127033,
+      "loss/hidden": 3.440234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18405662197619677,
+      "step": 18640
+    },
+    {
+      "epoch": 0.46625,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.8603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4266,
+      "loss/crossentropy": 2.194899543747306,
+      "loss/hidden": 3.3421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1942247728817165,
+      "step": 18650
+    },
+    {
+      "epoch": 0.4665,
+      "grad_norm": 28.25,
+      "grad_norm_var": 4.420572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2987,
+      "loss/crossentropy": 2.054854319989681,
+      "loss/hidden": 3.3203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17799932742491364,
+      "step": 18660
+    },
+    {
+      "epoch": 0.46675,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.7754557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3463,
+      "loss/crossentropy": 2.0972565457224848,
+      "loss/hidden": 3.3015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18171238731592892,
+      "step": 18670
+    },
+    {
+      "epoch": 0.467,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.254684908534436e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4225,
+      "loss/crossentropy": 2.2388961255550384,
+      "loss/hidden": 3.400390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20014150273054837,
+      "step": 18680
+    },
+    {
+      "epoch": 0.46725,
+      "grad_norm": 28.375,
+      "grad_norm_var": 8.96875,
+      "learning_rate": 0.0001,
+      "loss": 7.255,
+      "loss/crossentropy": 1.959997519850731,
+      "loss/hidden": 3.349609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1796635765582323,
+      "step": 18690
+    },
+    {
+      "epoch": 0.4675,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.253125,
+      "learning_rate": 0.0001,
+      "loss": 7.4076,
+      "loss/crossentropy": 1.9209066323935986,
+      "loss/hidden": 3.410546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18982454147189856,
+      "step": 18700
+    },
+    {
+      "epoch": 0.46775,
+      "grad_norm": 29.875,
+      "grad_norm_var": 4.195572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2657,
+      "loss/crossentropy": 2.0889955282211305,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17577018877491354,
+      "step": 18710
+    },
+    {
+      "epoch": 0.468,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.7997395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3168,
+      "loss/crossentropy": 2.0014477796852588,
+      "loss/hidden": 3.430078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18312612753361462,
+      "step": 18720
+    },
+    {
+      "epoch": 0.46825,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.222916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4403,
+      "loss/crossentropy": 2.1561698734760286,
+      "loss/hidden": 3.262109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17687523942440747,
+      "step": 18730
+    },
+    {
+      "epoch": 0.4685,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.4395182291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3355,
+      "loss/crossentropy": 2.116148530691862,
+      "loss/hidden": 3.330078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20347614474594594,
+      "step": 18740
+    },
+    {
+      "epoch": 0.46875,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.3020833333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3463,
+      "loss/crossentropy": 2.244252935051918,
+      "loss/hidden": 3.35625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19784990418702364,
+      "step": 18750
+    },
+    {
+      "epoch": 0.469,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.567122131243849e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2032,
+      "loss/crossentropy": 2.1832313030958175,
+      "loss/hidden": 3.41171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2140472937375307,
+      "step": 18760
+    },
+    {
+      "epoch": 0.46925,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.567122131550942e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.336,
+      "loss/crossentropy": 2.181801524013281,
+      "loss/hidden": 3.34609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18407561387866736,
+      "step": 18770
+    },
+    {
+      "epoch": 0.4695,
+      "grad_norm": 27.875,
+      "grad_norm_var": 1.78125,
+      "learning_rate": 0.0001,
+      "loss": 7.3641,
+      "loss/crossentropy": 2.1652345418930055,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19351193793118,
+      "step": 18780
+    },
+    {
+      "epoch": 0.46975,
+      "grad_norm": 28.25,
+      "grad_norm_var": 1.56875,
+      "learning_rate": 0.0001,
+      "loss": 7.2025,
+      "loss/crossentropy": 2.159587188065052,
+      "loss/hidden": 3.314453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18127773366868496,
+      "step": 18790
+    },
+    {
+      "epoch": 0.47,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.247916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3839,
+      "loss/crossentropy": 2.0050713911652567,
+      "loss/hidden": 3.420703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1919370399788022,
+      "step": 18800
+    },
+    {
+      "epoch": 0.47025,
+      "grad_norm": 34.25,
+      "grad_norm_var": 3.3997395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.34,
+      "loss/crossentropy": 2.121124693751335,
+      "loss/hidden": 3.3640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18033649511635302,
+      "step": 18810
+    },
+    {
+      "epoch": 0.4705,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.4650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.1365,
+      "loss/crossentropy": 1.9974480658769607,
+      "loss/hidden": 3.326953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17458155732601882,
+      "step": 18820
+    },
+    {
+      "epoch": 0.47075,
+      "grad_norm": 27.875,
+      "grad_norm_var": 2.5473307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3431,
+      "loss/crossentropy": 1.9829891242086888,
+      "loss/hidden": 3.37421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18825765140354633,
+      "step": 18830
+    },
+    {
+      "epoch": 0.471,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.7457682291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3446,
+      "loss/crossentropy": 2.1416624397039414,
+      "loss/hidden": 3.357421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19354026410728692,
+      "step": 18840
+    },
+    {
+      "epoch": 0.47125,
+      "grad_norm": 30.5,
+      "grad_norm_var": 11.280989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3904,
+      "loss/crossentropy": 2.072100041806698,
+      "loss/hidden": 3.383203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1879756365902722,
+      "step": 18850
+    },
+    {
+      "epoch": 0.4715,
+      "grad_norm": 28.5,
+      "grad_norm_var": 10.832291666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4158,
+      "loss/crossentropy": 2.196484336256981,
+      "loss/hidden": 3.383203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19391732811927795,
+      "step": 18860
+    },
+    {
+      "epoch": 0.47175,
+      "grad_norm": 29.0,
+      "grad_norm_var": 4.468489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.21,
+      "loss/crossentropy": 2.0025278739631176,
+      "loss/hidden": 3.26015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1782412089407444,
+      "step": 18870
+    },
+    {
+      "epoch": 0.472,
+      "grad_norm": 29.25,
+      "grad_norm_var": 2.85,
+      "learning_rate": 0.0001,
+      "loss": 7.2742,
+      "loss/crossentropy": 1.8140784852206706,
+      "loss/hidden": 3.375390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1594803395681083,
+      "step": 18880
+    },
+    {
+      "epoch": 0.47225,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.594153947490541e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4099,
+      "loss/crossentropy": 2.0685198314487936,
+      "loss/hidden": 3.35859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17976643573492765,
+      "step": 18890
+    },
+    {
+      "epoch": 0.4725,
+      "grad_norm": 27.5,
+      "grad_norm_var": 3.5941539466769157e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3278,
+      "loss/crossentropy": 1.995526134967804,
+      "loss/hidden": 3.41484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20469516087323428,
+      "step": 18900
+    },
+    {
+      "epoch": 0.47275,
+      "grad_norm": 28.5,
+      "grad_norm_var": 3.9176432291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2844,
+      "loss/crossentropy": 2.081358040869236,
+      "loss/hidden": 3.293359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1750167841091752,
+      "step": 18910
+    },
+    {
+      "epoch": 0.473,
+      "grad_norm": 28.25,
+      "grad_norm_var": 1.0518229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2403,
+      "loss/crossentropy": 2.08730803206563,
+      "loss/hidden": 3.347265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1881487349048257,
+      "step": 18920
+    },
+    {
+      "epoch": 0.47325,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.4871128695672494e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3072,
+      "loss/crossentropy": 2.050686779618263,
+      "loss/hidden": 3.263671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18110219333320857,
+      "step": 18930
+    },
+    {
+      "epoch": 0.4735,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.48711286963296e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3354,
+      "loss/crossentropy": 2.116658928990364,
+      "loss/hidden": 3.318359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18662892282009125,
+      "step": 18940
+    },
+    {
+      "epoch": 0.47375,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.417122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2869,
+      "loss/crossentropy": 2.181796707212925,
+      "loss/hidden": 3.31171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1841755596920848,
+      "step": 18950
+    },
+    {
+      "epoch": 0.474,
+      "grad_norm": 31.375,
+      "grad_norm_var": 4.01640625,
+      "learning_rate": 0.0001,
+      "loss": 7.238,
+      "loss/crossentropy": 2.096955654025078,
+      "loss/hidden": 3.31015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1859470222145319,
+      "step": 18960
+    },
+    {
+      "epoch": 0.47425,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.2020833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2649,
+      "loss/crossentropy": 2.1101110115647317,
+      "loss/hidden": 3.2515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18113658875226973,
+      "step": 18970
+    },
+    {
+      "epoch": 0.4745,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.6952473958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.218,
+      "loss/crossentropy": 1.9812964178621768,
+      "loss/hidden": 3.343359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17663604663684965,
+      "step": 18980
+    },
+    {
+      "epoch": 0.47475,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.8926432291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.464,
+      "loss/crossentropy": 2.2813973009586332,
+      "loss/hidden": 3.28203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19019655492156745,
+      "step": 18990
+    },
+    {
+      "epoch": 0.475,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.423958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.427,
+      "loss/crossentropy": 2.1718254670500756,
+      "loss/hidden": 3.2765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18812184669077398,
+      "step": 19000
+    },
+    {
+      "epoch": 0.47525,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.986168000216896e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4384,
+      "loss/crossentropy": 1.9550217375159264,
+      "loss/hidden": 3.39140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17679040897637605,
+      "step": 19010
+    },
+    {
+      "epoch": 0.4755,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.9861680010449213e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3947,
+      "loss/crossentropy": 2.001023256778717,
+      "loss/hidden": 3.3453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18076751604676247,
+      "step": 19020
+    },
+    {
+      "epoch": 0.47575,
+      "grad_norm": 27.375,
+      "grad_norm_var": 1.3268229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2484,
+      "loss/crossentropy": 2.0497733250260355,
+      "loss/hidden": 3.19453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1572325760498643,
+      "step": 19030
+    },
+    {
+      "epoch": 0.476,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.796875,
+      "learning_rate": 0.0001,
+      "loss": 7.4071,
+      "loss/crossentropy": 2.1493683993816375,
+      "loss/hidden": 3.330859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19892700631171464,
+      "step": 19040
+    },
+    {
+      "epoch": 0.47625,
+      "grad_norm": 35.0,
+      "grad_norm_var": 3.356705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2913,
+      "loss/crossentropy": 2.0519852221012114,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1828221082687378,
+      "step": 19050
+    },
+    {
+      "epoch": 0.4765,
+      "grad_norm": 36.0,
+      "grad_norm_var": 3.5624176427021107e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3334,
+      "loss/crossentropy": 2.201715832948685,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20039877742528917,
+      "step": 19060
+    },
+    {
+      "epoch": 0.47675,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.562417643000955e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3979,
+      "loss/crossentropy": 2.0442496329545974,
+      "loss/hidden": 3.432421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1853096466511488,
+      "step": 19070
+    },
+    {
+      "epoch": 0.477,
+      "grad_norm": 30.0,
+      "grad_norm_var": 1.5322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2213,
+      "loss/crossentropy": 2.0714069202542307,
+      "loss/hidden": 3.265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1765752835199237,
+      "step": 19080
+    },
+    {
+      "epoch": 0.47725,
+      "grad_norm": 28.0,
+      "grad_norm_var": 2.2462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2175,
+      "loss/crossentropy": 1.784161239117384,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19739716919139028,
+      "step": 19090
+    },
+    {
+      "epoch": 0.4775,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.033072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.258,
+      "loss/crossentropy": 2.0708198331296446,
+      "loss/hidden": 3.43515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18721713982522487,
+      "step": 19100
+    },
+    {
+      "epoch": 0.47775,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.3442041695808343e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.363,
+      "loss/crossentropy": 2.105399575829506,
+      "loss/hidden": 3.39296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19771641138941048,
+      "step": 19110
+    },
+    {
+      "epoch": 0.478,
+      "grad_norm": 31.25,
+      "grad_norm_var": 6.77890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3457,
+      "loss/crossentropy": 2.0999079287052154,
+      "loss/hidden": 3.265234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1782015733420849,
+      "step": 19120
+    },
+    {
+      "epoch": 0.47825,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.011458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2927,
+      "loss/crossentropy": 2.060526109486818,
+      "loss/hidden": 3.454296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19682135116308927,
+      "step": 19130
+    },
+    {
+      "epoch": 0.4785,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.3447265625,
+      "learning_rate": 0.0001,
+      "loss": 7.1563,
+      "loss/crossentropy": 1.9666556693613528,
+      "loss/hidden": 3.39375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18265628404915332,
+      "step": 19140
+    },
+    {
+      "epoch": 0.47875,
+      "grad_norm": 29.0,
+      "grad_norm_var": 1.0858723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3782,
+      "loss/crossentropy": 2.233207347989082,
+      "loss/hidden": 3.305859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17987454514950513,
+      "step": 19150
+    },
+    {
+      "epoch": 0.479,
+      "grad_norm": 30.75,
+      "grad_norm_var": 32.65807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2859,
+      "loss/crossentropy": 2.304356482625008,
+      "loss/hidden": 3.309375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19401859547942876,
+      "step": 19160
+    },
+    {
+      "epoch": 0.47925,
+      "grad_norm": 26.75,
+      "grad_norm_var": 35.8134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.1512,
+      "loss/crossentropy": 2.187753638625145,
+      "loss/hidden": 3.326953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19042326286435127,
+      "step": 19170
+    },
+    {
+      "epoch": 0.4795,
+      "grad_norm": 27.875,
+      "grad_norm_var": 3.17890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3497,
+      "loss/crossentropy": 2.25252815335989,
+      "loss/hidden": 3.34453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1958579957485199,
+      "step": 19180
+    },
+    {
+      "epoch": 0.47975,
+      "grad_norm": 33.25,
+      "grad_norm_var": 3.4058471885172465e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2983,
+      "loss/crossentropy": 1.9786495432257651,
+      "loss/hidden": 3.376953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19189926832914353,
+      "step": 19190
+    },
+    {
+      "epoch": 0.48,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.405847187102368e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2884,
+      "loss/crossentropy": 2.13499116525054,
+      "loss/hidden": 3.2828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18915282813832163,
+      "step": 19200
+    },
+    {
+      "epoch": 0.48025,
+      "grad_norm": 29.375,
+      "grad_norm_var": 7.429622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3051,
+      "loss/crossentropy": 2.1838356882333754,
+      "loss/hidden": 3.423828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19484667405486106,
+      "step": 19210
+    },
+    {
+      "epoch": 0.4805,
+      "grad_norm": 35.0,
+      "grad_norm_var": 5.877018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2957,
+      "loss/crossentropy": 2.1596252396702766,
+      "loss/hidden": 3.422265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19484965018928052,
+      "step": 19220
+    },
+    {
+      "epoch": 0.48075,
+      "grad_norm": 32.25,
+      "grad_norm_var": 6.5744140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3458,
+      "loss/crossentropy": 1.9286774158477784,
+      "loss/hidden": 3.455859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20619444027543068,
+      "step": 19230
+    },
+    {
+      "epoch": 0.481,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.6780598958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3249,
+      "loss/crossentropy": 2.0902815729379656,
+      "loss/hidden": 3.285546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17610749285668134,
+      "step": 19240
+    },
+    {
+      "epoch": 0.48125,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.191666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3046,
+      "loss/crossentropy": 2.117246875911951,
+      "loss/hidden": 3.3140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1708180472254753,
+      "step": 19250
+    },
+    {
+      "epoch": 0.4815,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.9801432291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.285,
+      "loss/crossentropy": 1.9743897125124932,
+      "loss/hidden": 3.475,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18220343571156264,
+      "step": 19260
+    },
+    {
+      "epoch": 0.48175,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.52265625,
+      "learning_rate": 0.0001,
+      "loss": 7.416,
+      "loss/crossentropy": 2.128947339951992,
+      "loss/hidden": 3.32890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18609113954007625,
+      "step": 19270
+    },
+    {
+      "epoch": 0.482,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.0311848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4237,
+      "loss/crossentropy": 1.9989797204732895,
+      "loss/hidden": 3.436328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19045947063714266,
+      "step": 19280
+    },
+    {
+      "epoch": 0.48225,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.456184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3041,
+      "loss/crossentropy": 2.0317627035081385,
+      "loss/hidden": 3.33984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18456566017121076,
+      "step": 19290
+    },
+    {
+      "epoch": 0.4825,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.0416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2606,
+      "loss/crossentropy": 1.994530802220106,
+      "loss/hidden": 3.36328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18545446768403054,
+      "step": 19300
+    },
+    {
+      "epoch": 0.48275,
+      "grad_norm": 30.75,
+      "grad_norm_var": 8.439518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.352,
+      "loss/crossentropy": 2.023553788661957,
+      "loss/hidden": 3.30703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17272286228835582,
+      "step": 19310
+    },
+    {
+      "epoch": 0.483,
+      "grad_norm": 27.625,
+      "grad_norm_var": 8.5462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3492,
+      "loss/crossentropy": 2.0078469097614287,
+      "loss/hidden": 3.45078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1847759123891592,
+      "step": 19320
+    },
+    {
+      "epoch": 0.48325,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.8020833333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2202,
+      "loss/crossentropy": 2.084314952790737,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17277139481157064,
+      "step": 19330
+    },
+    {
+      "epoch": 0.4835,
+      "grad_norm": 27.875,
+      "grad_norm_var": 2.183333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3691,
+      "loss/crossentropy": 2.0023327678442002,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17878497913479804,
+      "step": 19340
+    },
+    {
+      "epoch": 0.48375,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.9556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2622,
+      "loss/crossentropy": 2.160814478993416,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19914763923734427,
+      "step": 19350
+    },
+    {
+      "epoch": 0.484,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.22265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4019,
+      "loss/crossentropy": 2.0562733739614485,
+      "loss/hidden": 3.521484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.231661388091743,
+      "step": 19360
+    },
+    {
+      "epoch": 0.48425,
+      "grad_norm": 32.5,
+      "grad_norm_var": 5.6650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3196,
+      "loss/crossentropy": 2.1037007592618466,
+      "loss/hidden": 3.41328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20012609604746104,
+      "step": 19370
+    },
+    {
+      "epoch": 0.4845,
+      "grad_norm": 30.25,
+      "grad_norm_var": 5.585872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2967,
+      "loss/crossentropy": 2.0158598870038986,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18958695326000452,
+      "step": 19380
+    },
+    {
+      "epoch": 0.48475,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.6494140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3039,
+      "loss/crossentropy": 2.155204123258591,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2042725382372737,
+      "step": 19390
+    },
+    {
+      "epoch": 0.485,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.001822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3471,
+      "loss/crossentropy": 2.130756896734238,
+      "loss/hidden": 3.48984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21289011891931295,
+      "step": 19400
+    },
+    {
+      "epoch": 0.48525,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.2583333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3385,
+      "loss/crossentropy": 1.9661409441381692,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1758869204670191,
+      "step": 19410
+    },
+    {
+      "epoch": 0.4855,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.2416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3636,
+      "loss/crossentropy": 2.088372728228569,
+      "loss/hidden": 3.320703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1817894596606493,
+      "step": 19420
+    },
+    {
+      "epoch": 0.48575,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.9983723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2972,
+      "loss/crossentropy": 2.288235005736351,
+      "loss/hidden": 3.339453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20123199094086885,
+      "step": 19430
+    },
+    {
+      "epoch": 0.486,
+      "grad_norm": 30.25,
+      "grad_norm_var": 0.8858723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3093,
+      "loss/crossentropy": 2.1147582471370696,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1947942255064845,
+      "step": 19440
+    },
+    {
+      "epoch": 0.48625,
+      "grad_norm": 34.0,
+      "grad_norm_var": 2.8509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3116,
+      "loss/crossentropy": 2.089379907399416,
+      "loss/hidden": 3.38125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19097770117223262,
+      "step": 19450
+    },
+    {
+      "epoch": 0.4865,
+      "grad_norm": 29.25,
+      "grad_norm_var": 2.7372395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3315,
+      "loss/crossentropy": 2.16123593300581,
+      "loss/hidden": 3.43359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19270185381174088,
+      "step": 19460
+    },
+    {
+      "epoch": 0.48675,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.511458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3166,
+      "loss/crossentropy": 1.9626320779323578,
+      "loss/hidden": 3.41640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1842844920232892,
+      "step": 19470
+    },
+    {
+      "epoch": 0.487,
+      "grad_norm": 28.625,
+      "grad_norm_var": 2.5254557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3752,
+      "loss/crossentropy": 2.0067977510392665,
+      "loss/hidden": 3.446484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1906943406909704,
+      "step": 19480
+    },
+    {
+      "epoch": 0.48725,
+      "grad_norm": 30.25,
+      "grad_norm_var": 5.573372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3081,
+      "loss/crossentropy": 1.9796999536454678,
+      "loss/hidden": 3.3046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17342409659177066,
+      "step": 19490
+    },
+    {
+      "epoch": 0.4875,
+      "grad_norm": 31.25,
+      "grad_norm_var": 6.215559895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2416,
+      "loss/crossentropy": 2.0168533116579055,
+      "loss/hidden": 3.288671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17710478082299233,
+      "step": 19500
+    },
+    {
+      "epoch": 0.48775,
+      "grad_norm": 28.375,
+      "grad_norm_var": 5.944791666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2389,
+      "loss/crossentropy": 2.0261049672961233,
+      "loss/hidden": 3.309765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17297806199640037,
+      "step": 19510
+    },
+    {
+      "epoch": 0.488,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.620247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3401,
+      "loss/crossentropy": 2.156434553861618,
+      "loss/hidden": 3.345703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19226507674902676,
+      "step": 19520
+    },
+    {
+      "epoch": 0.48825,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.568489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2962,
+      "loss/crossentropy": 2.0590639621019364,
+      "loss/hidden": 3.277734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17470778487622737,
+      "step": 19530
+    },
+    {
+      "epoch": 0.4885,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.6957682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2644,
+      "loss/crossentropy": 2.2576445043087006,
+      "loss/hidden": 3.30390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17541414052248,
+      "step": 19540
+    },
+    {
+      "epoch": 0.48875,
+      "grad_norm": 33.25,
+      "grad_norm_var": 5.232747395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2863,
+      "loss/crossentropy": 1.8561327636241913,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17461737468838692,
+      "step": 19550
+    },
+    {
+      "epoch": 0.489,
+      "grad_norm": 30.25,
+      "grad_norm_var": 9.878125,
+      "learning_rate": 0.0001,
+      "loss": 7.2798,
+      "loss/crossentropy": 1.989323940873146,
+      "loss/hidden": 3.3515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19638592712581157,
+      "step": 19560
+    },
+    {
+      "epoch": 0.48925,
+      "grad_norm": 32.5,
+      "grad_norm_var": 13.573893229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4564,
+      "loss/crossentropy": 2.173425105214119,
+      "loss/hidden": 3.38828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19522506278008223,
+      "step": 19570
+    },
+    {
+      "epoch": 0.4895,
+      "grad_norm": 32.25,
+      "grad_norm_var": 13.3087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4398,
+      "loss/crossentropy": 2.1963840425014496,
+      "loss/hidden": 3.383203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20581881813704966,
+      "step": 19580
+    },
+    {
+      "epoch": 0.48975,
+      "grad_norm": 31.375,
+      "grad_norm_var": 14.501497395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.288,
+      "loss/crossentropy": 2.2206703200936317,
+      "loss/hidden": 3.310546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1854398850351572,
+      "step": 19590
+    },
+    {
+      "epoch": 0.49,
+      "grad_norm": 29.125,
+      "grad_norm_var": 6.174934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3071,
+      "loss/crossentropy": 2.182942679524422,
+      "loss/hidden": 3.448046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20937685370445253,
+      "step": 19600
+    },
+    {
+      "epoch": 0.49025,
+      "grad_norm": 30.625,
+      "grad_norm_var": 5.620768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2791,
+      "loss/crossentropy": 1.968775314092636,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18280262742191553,
+      "step": 19610
+    },
+    {
+      "epoch": 0.4905,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.764322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2851,
+      "loss/crossentropy": 2.1555724523961546,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1888716220855713,
+      "step": 19620
+    },
+    {
+      "epoch": 0.49075,
+      "grad_norm": 36.5,
+      "grad_norm_var": 7.4150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2932,
+      "loss/crossentropy": 2.2111911468207834,
+      "loss/hidden": 3.321875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18723924830555916,
+      "step": 19630
+    },
+    {
+      "epoch": 0.491,
+      "grad_norm": 31.25,
+      "grad_norm_var": 6.720572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2567,
+      "loss/crossentropy": 1.9819117814302445,
+      "loss/hidden": 3.3390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17056979294866323,
+      "step": 19640
+    },
+    {
+      "epoch": 0.49125,
+      "grad_norm": 28.0,
+      "grad_norm_var": 3.317643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2356,
+      "loss/crossentropy": 2.142751504480839,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17397481221705674,
+      "step": 19650
+    },
+    {
+      "epoch": 0.4915,
+      "grad_norm": 31.125,
+      "grad_norm_var": 5.001041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3434,
+      "loss/crossentropy": 2.073681902885437,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2028063558973372,
+      "step": 19660
+    },
+    {
+      "epoch": 0.49175,
+      "grad_norm": 29.375,
+      "grad_norm_var": 4.9697265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2295,
+      "loss/crossentropy": 2.070267468690872,
+      "loss/hidden": 3.421484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18656076975166797,
+      "step": 19670
+    },
+    {
+      "epoch": 0.492,
+      "grad_norm": 32.25,
+      "grad_norm_var": 6.004166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3869,
+      "loss/crossentropy": 2.0801706589758395,
+      "loss/hidden": 3.477734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2120389549061656,
+      "step": 19680
+    },
+    {
+      "epoch": 0.49225,
+      "grad_norm": 35.0,
+      "grad_norm_var": 4.446809895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4227,
+      "loss/crossentropy": 2.1620673507452013,
+      "loss/hidden": 3.35078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.197611235268414,
+      "step": 19690
+    },
+    {
+      "epoch": 0.4925,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.01640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3952,
+      "loss/crossentropy": 2.102504736185074,
+      "loss/hidden": 3.260546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18240329679101705,
+      "step": 19700
+    },
+    {
+      "epoch": 0.49275,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.0444333199307136e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2426,
+      "loss/crossentropy": 2.2651669278740885,
+      "loss/hidden": 3.295703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18710593543946744,
+      "step": 19710
+    },
+    {
+      "epoch": 0.493,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.044433320824939e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2411,
+      "loss/crossentropy": 2.185315527021885,
+      "loss/hidden": 3.33203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1886447848752141,
+      "step": 19720
+    },
+    {
+      "epoch": 0.49325,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.78125,
+      "learning_rate": 0.0001,
+      "loss": 7.3084,
+      "loss/crossentropy": 2.066023611277342,
+      "loss/hidden": 3.35703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1871557403355837,
+      "step": 19730
+    },
+    {
+      "epoch": 0.4935,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.215625,
+      "learning_rate": 0.0001,
+      "loss": 7.2056,
+      "loss/crossentropy": 2.2202082961797713,
+      "loss/hidden": 3.274609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17992934715002776,
+      "step": 19740
+    },
+    {
+      "epoch": 0.49375,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.4452473958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3547,
+      "loss/crossentropy": 2.0926157012581825,
+      "loss/hidden": 3.322265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19089100752025842,
+      "step": 19750
+    },
+    {
+      "epoch": 0.494,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.0973307291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3392,
+      "loss/crossentropy": 2.0247383177280427,
+      "loss/hidden": 3.434765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19218139182776212,
+      "step": 19760
+    },
+    {
+      "epoch": 0.49425,
+      "grad_norm": 6677331968.0,
+      "grad_norm_var": 2.786672612235455e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3587,
+      "loss/crossentropy": 1.9743504285812379,
+      "loss/hidden": 3.4015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1963216718286276,
+      "step": 19770
+    },
+    {
+      "epoch": 0.4945,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.7866726115468554e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3528,
+      "loss/crossentropy": 2.065458830446005,
+      "loss/hidden": 3.327734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1823502728715539,
+      "step": 19780
+    },
+    {
+      "epoch": 0.49475,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.9124348958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3247,
+      "loss/crossentropy": 2.108359482139349,
+      "loss/hidden": 3.42734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2117020718753338,
+      "step": 19790
+    },
+    {
+      "epoch": 0.495,
+      "grad_norm": 28.375,
+      "grad_norm_var": 3.7059895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3429,
+      "loss/crossentropy": 2.245875895023346,
+      "loss/hidden": 3.28046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1797666964121163,
+      "step": 19800
+    },
+    {
+      "epoch": 0.49525,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.463997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3058,
+      "loss/crossentropy": 2.031013163924217,
+      "loss/hidden": 3.34453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17521858643740415,
+      "step": 19810
+    },
+    {
+      "epoch": 0.4955,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.26875,
+      "learning_rate": 0.0001,
+      "loss": 7.2214,
+      "loss/crossentropy": 2.048213458806276,
+      "loss/hidden": 3.441796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1798891793936491,
+      "step": 19820
+    },
+    {
+      "epoch": 0.49575,
+      "grad_norm": 30.5,
+      "grad_norm_var": 18.353059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3332,
+      "loss/crossentropy": 2.1515459649264814,
+      "loss/hidden": 3.348828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1860925205051899,
+      "step": 19830
+    },
+    {
+      "epoch": 0.496,
+      "grad_norm": 29.125,
+      "grad_norm_var": 19.2275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3004,
+      "loss/crossentropy": 2.1721823632717134,
+      "loss/hidden": 3.360546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19930550809949638,
+      "step": 19840
+    },
+    {
+      "epoch": 0.49625,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.5072265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2698,
+      "loss/crossentropy": 1.9935913749039174,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17680493406951428,
+      "step": 19850
+    },
+    {
+      "epoch": 0.4965,
+      "grad_norm": 6945767424.0,
+      "grad_norm_var": 3.015230293460858e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2426,
+      "loss/crossentropy": 2.069664215296507,
+      "loss/hidden": 3.569921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18678983384743333,
+      "step": 19860
+    },
+    {
+      "epoch": 0.49675,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.015230291984882e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3594,
+      "loss/crossentropy": 2.0391472943127154,
+      "loss/hidden": 3.392578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18784078639000654,
+      "step": 19870
+    },
+    {
+      "epoch": 0.497,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.3822916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2527,
+      "loss/crossentropy": 2.046357312053442,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17729539722204207,
+      "step": 19880
+    },
+    {
+      "epoch": 0.49725,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.7872395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.307,
+      "loss/crossentropy": 2.021935646235943,
+      "loss/hidden": 3.46015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20562903992831708,
+      "step": 19890
+    },
+    {
+      "epoch": 0.4975,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.1747395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3574,
+      "loss/crossentropy": 2.0218880496919156,
+      "loss/hidden": 3.408203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1812274256721139,
+      "step": 19900
+    },
+    {
+      "epoch": 0.49775,
+      "grad_norm": 42.5,
+      "grad_norm_var": 11.66640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2552,
+      "loss/crossentropy": 2.1163423866033555,
+      "loss/hidden": 3.3015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18092286475002767,
+      "step": 19910
+    },
+    {
+      "epoch": 0.498,
+      "grad_norm": 31.0,
+      "grad_norm_var": 13.023893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4151,
+      "loss/crossentropy": 2.25559261739254,
+      "loss/hidden": 3.35625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19431090019643307,
+      "step": 19920
+    },
+    {
+      "epoch": 0.49825,
+      "grad_norm": 34.0,
+      "grad_norm_var": 9.103059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.33,
+      "loss/crossentropy": 2.1288936853408815,
+      "loss/hidden": 3.2796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.185909984074533,
+      "step": 19930
+    },
+    {
+      "epoch": 0.4985,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.645768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3463,
+      "loss/crossentropy": 2.1082677975296975,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18483771663159132,
+      "step": 19940
+    },
+    {
+      "epoch": 0.49875,
+      "grad_norm": 29.625,
+      "grad_norm_var": 1.3114583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3146,
+      "loss/crossentropy": 2.047964480519295,
+      "loss/hidden": 3.337109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18684354815632104,
+      "step": 19950
+    },
+    {
+      "epoch": 0.499,
+      "grad_norm": 30.25,
+      "grad_norm_var": 7.695833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1954,
+      "loss/crossentropy": 2.0788462795317173,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18402842283248902,
+      "step": 19960
+    },
+    {
+      "epoch": 0.49925,
+      "grad_norm": 31.375,
+      "grad_norm_var": 7.2056640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3153,
+      "loss/crossentropy": 1.910679142177105,
+      "loss/hidden": 3.321875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16827462129294873,
+      "step": 19970
+    },
+    {
+      "epoch": 0.4995,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.7434895833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2924,
+      "loss/crossentropy": 2.133229525387287,
+      "loss/hidden": 3.323046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18238822668790816,
+      "step": 19980
+    },
+    {
+      "epoch": 0.49975,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.7212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3936,
+      "loss/crossentropy": 2.1271730199456216,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19191278666257858,
+      "step": 19990
+    },
+    {
+      "epoch": 0.5,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.5551432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2894,
+      "loss/crossentropy": 2.0791160173714163,
+      "loss/hidden": 3.356640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1787281386554241,
+      "step": 20000
+    },
+    {
+      "epoch": 0.50025,
+      "grad_norm": 30.125,
+      "grad_norm_var": 4.286393229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3934,
+      "loss/crossentropy": 2.100822980701923,
+      "loss/hidden": 3.31640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1837709965184331,
+      "step": 20010
+    },
+    {
+      "epoch": 0.5005,
+      "grad_norm": 27.0,
+      "grad_norm_var": 6.8275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2903,
+      "loss/crossentropy": 2.0603928424417974,
+      "loss/hidden": 3.3328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18236217414960265,
+      "step": 20020
+    },
+    {
+      "epoch": 0.50075,
+      "grad_norm": 31.125,
+      "grad_norm_var": 5.8509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2986,
+      "loss/crossentropy": 2.0794493168592454,
+      "loss/hidden": 3.341796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17758788168430328,
+      "step": 20030
+    },
+    {
+      "epoch": 0.501,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.5884765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3217,
+      "loss/crossentropy": 2.180893415212631,
+      "loss/hidden": 3.390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19812680166214705,
+      "step": 20040
+    },
+    {
+      "epoch": 0.50125,
+      "grad_norm": 29.25,
+      "grad_norm_var": 2.2718098958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2937,
+      "loss/crossentropy": 2.0710356786847113,
+      "loss/hidden": 3.359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18272985238581896,
+      "step": 20050
+    },
+    {
+      "epoch": 0.5015,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.61015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3658,
+      "loss/crossentropy": 2.2894588097929955,
+      "loss/hidden": 3.30859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19785721227526665,
+      "step": 20060
+    },
+    {
+      "epoch": 0.50175,
+      "grad_norm": 33.0,
+      "grad_norm_var": 6.8572265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3323,
+      "loss/crossentropy": 1.9864086501300335,
+      "loss/hidden": 3.41015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18656586892902852,
+      "step": 20070
+    },
+    {
+      "epoch": 0.502,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.0302083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2853,
+      "loss/crossentropy": 2.08188853263855,
+      "loss/hidden": 3.347265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1811662282794714,
+      "step": 20080
+    },
+    {
+      "epoch": 0.50225,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.2728515625,
+      "learning_rate": 0.0001,
+      "loss": 7.1962,
+      "loss/crossentropy": 2.175425173342228,
+      "loss/hidden": 3.341796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18717424534261226,
+      "step": 20090
+    },
+    {
+      "epoch": 0.5025,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.0770833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2638,
+      "loss/crossentropy": 2.1691887110471724,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19185122177004815,
+      "step": 20100
+    },
+    {
+      "epoch": 0.50275,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.1285807291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2232,
+      "loss/crossentropy": 2.2031719744205476,
+      "loss/hidden": 3.280859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18911925759166479,
+      "step": 20110
+    },
+    {
+      "epoch": 0.503,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.6791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.344,
+      "loss/crossentropy": 1.9854541048407555,
+      "loss/hidden": 3.321484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17294510118663312,
+      "step": 20120
+    },
+    {
+      "epoch": 0.50325,
+      "grad_norm": 29.25,
+      "grad_norm_var": 1.9504557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.291,
+      "loss/crossentropy": 2.098982587456703,
+      "loss/hidden": 3.283203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18347709700465203,
+      "step": 20130
+    },
+    {
+      "epoch": 0.5035,
+      "grad_norm": 28.5,
+      "grad_norm_var": 8.94765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3558,
+      "loss/crossentropy": 2.0379406858235596,
+      "loss/hidden": 3.4921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20050731115043163,
+      "step": 20140
+    },
+    {
+      "epoch": 0.50375,
+      "grad_norm": 29.0,
+      "grad_norm_var": 10.076822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4874,
+      "loss/crossentropy": 2.1478746205568315,
+      "loss/hidden": 3.316796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19442609827965499,
+      "step": 20150
+    },
+    {
+      "epoch": 0.504,
+      "grad_norm": 29.125,
+      "grad_norm_var": 18.79765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2431,
+      "loss/crossentropy": 2.0215858951210977,
+      "loss/hidden": 3.337109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1762773571535945,
+      "step": 20160
+    },
+    {
+      "epoch": 0.50425,
+      "grad_norm": 31.75,
+      "grad_norm_var": 19.851041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2569,
+      "loss/crossentropy": 2.053553320467472,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18032009098678828,
+      "step": 20170
+    },
+    {
+      "epoch": 0.5045,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.47265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3017,
+      "loss/crossentropy": 2.1755366683006288,
+      "loss/hidden": 3.36015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19120553210377694,
+      "step": 20180
+    },
+    {
+      "epoch": 0.50475,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.2145182291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2499,
+      "loss/crossentropy": 2.0905803576111794,
+      "loss/hidden": 3.333984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18027825355529786,
+      "step": 20190
+    },
+    {
+      "epoch": 0.505,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.0837890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3299,
+      "loss/crossentropy": 2.1392801135778425,
+      "loss/hidden": 3.40234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18984563760459422,
+      "step": 20200
+    },
+    {
+      "epoch": 0.50525,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.7705729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3683,
+      "loss/crossentropy": 1.9904960945248604,
+      "loss/hidden": 3.34765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19267166014760734,
+      "step": 20210
+    },
+    {
+      "epoch": 0.5055,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.4369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4132,
+      "loss/crossentropy": 2.118674224615097,
+      "loss/hidden": 3.265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18962926007807254,
+      "step": 20220
+    },
+    {
+      "epoch": 0.50575,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.5275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2848,
+      "loss/crossentropy": 2.042106767743826,
+      "loss/hidden": 3.420703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1888286828994751,
+      "step": 20230
+    },
+    {
+      "epoch": 0.506,
+      "grad_norm": 43.5,
+      "grad_norm_var": 11.646875,
+      "learning_rate": 0.0001,
+      "loss": 7.3466,
+      "loss/crossentropy": 2.01454746350646,
+      "loss/hidden": 3.33984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17999137863516806,
+      "step": 20240
+    },
+    {
+      "epoch": 0.50625,
+      "grad_norm": 30.625,
+      "grad_norm_var": 16.970572916666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3463,
+      "loss/crossentropy": 2.0871743500232696,
+      "loss/hidden": 3.298046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18978464882820845,
+      "step": 20250
+    },
+    {
+      "epoch": 0.5065,
+      "grad_norm": 29.75,
+      "grad_norm_var": 23.7556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3989,
+      "loss/crossentropy": 2.0043682172894477,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20130981262773276,
+      "step": 20260
+    },
+    {
+      "epoch": 0.50675,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.066666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3753,
+      "loss/crossentropy": 2.0674937814474106,
+      "loss/hidden": 3.43046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1961617423221469,
+      "step": 20270
+    },
+    {
+      "epoch": 0.507,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.8934895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2008,
+      "loss/crossentropy": 2.1054218977689745,
+      "loss/hidden": 3.321875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1805409086868167,
+      "step": 20280
+    },
+    {
+      "epoch": 0.50725,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.3395182291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3738,
+      "loss/crossentropy": 2.1355152033269404,
+      "loss/hidden": 3.37578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18312898697331548,
+      "step": 20290
+    },
+    {
+      "epoch": 0.5075,
+      "grad_norm": 44.5,
+      "grad_norm_var": 14.501041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3439,
+      "loss/crossentropy": 1.953713247179985,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16738616386428476,
+      "step": 20300
+    },
+    {
+      "epoch": 0.50775,
+      "grad_norm": 30.0,
+      "grad_norm_var": 13.852083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3785,
+      "loss/crossentropy": 2.1630780398845673,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18321501482278107,
+      "step": 20310
+    },
+    {
+      "epoch": 0.508,
+      "grad_norm": 35.25,
+      "grad_norm_var": 3.4587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3335,
+      "loss/crossentropy": 1.973591250926256,
+      "loss/hidden": 3.419140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1923765016719699,
+      "step": 20320
+    },
+    {
+      "epoch": 0.50825,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2324,
+      "loss/crossentropy": 1.98987924605608,
+      "loss/hidden": 3.362109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18974843434989452,
+      "step": 20330
+    },
+    {
+      "epoch": 0.5085,
+      "grad_norm": 28.75,
+      "grad_norm_var": 6.96875,
+      "learning_rate": 0.0001,
+      "loss": 7.3109,
+      "loss/crossentropy": 2.277467969059944,
+      "loss/hidden": 3.317578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19834709204733372,
+      "step": 20340
+    },
+    {
+      "epoch": 0.50875,
+      "grad_norm": 29.625,
+      "grad_norm_var": 11.067122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2231,
+      "loss/crossentropy": 1.959492462873459,
+      "loss/hidden": 3.41796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1740871286019683,
+      "step": 20350
+    },
+    {
+      "epoch": 0.509,
+      "grad_norm": 38.0,
+      "grad_norm_var": 8.055989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3223,
+      "loss/crossentropy": 2.1561495400965214,
+      "loss/hidden": 3.3921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18467604517936706,
+      "step": 20360
+    },
+    {
+      "epoch": 0.50925,
+      "grad_norm": 36.25,
+      "grad_norm_var": 6.133333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3836,
+      "loss/crossentropy": 2.0499372258782387,
+      "loss/hidden": 3.3078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17811761405318977,
+      "step": 20370
+    },
+    {
+      "epoch": 0.5095,
+      "grad_norm": 36.75,
+      "grad_norm_var": 6.571809895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3418,
+      "loss/crossentropy": 2.0755091689527037,
+      "loss/hidden": 3.536328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2004254411906004,
+      "step": 20380
+    },
+    {
+      "epoch": 0.50975,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.4083702435510134e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4065,
+      "loss/crossentropy": 2.1109745904803274,
+      "loss/hidden": 3.345703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18784465081989765,
+      "step": 20390
+    },
+    {
+      "epoch": 0.51,
+      "grad_norm": 34.25,
+      "grad_norm_var": 2.408370242716871e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3957,
+      "loss/crossentropy": 2.204335790872574,
+      "loss/hidden": 3.301171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1780173137784004,
+      "step": 20400
+    },
+    {
+      "epoch": 0.51025,
+      "grad_norm": 32.0,
+      "grad_norm_var": 41.43723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1806,
+      "loss/crossentropy": 1.9774738259613514,
+      "loss/hidden": 3.417578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18556309957057238,
+      "step": 20410
+    },
+    {
+      "epoch": 0.5105,
+      "grad_norm": 27.625,
+      "grad_norm_var": 65.22473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3462,
+      "loss/crossentropy": 2.167100024223328,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18838558737188577,
+      "step": 20420
+    },
+    {
+      "epoch": 0.51075,
+      "grad_norm": 28.375,
+      "grad_norm_var": 36.87291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3272,
+      "loss/crossentropy": 2.005172957479954,
+      "loss/hidden": 3.358203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19127585394307972,
+      "step": 20430
+    },
+    {
+      "epoch": 0.511,
+      "grad_norm": 30.25,
+      "grad_norm_var": 37.78125,
+      "learning_rate": 0.0001,
+      "loss": 7.226,
+      "loss/crossentropy": 2.063726345449686,
+      "loss/hidden": 3.424609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19834220120683313,
+      "step": 20440
+    },
+    {
+      "epoch": 0.51125,
+      "grad_norm": 35.0,
+      "grad_norm_var": 12.229166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2767,
+      "loss/crossentropy": 2.0135110549628736,
+      "loss/hidden": 3.295703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17450594212859868,
+      "step": 20450
+    },
+    {
+      "epoch": 0.5115,
+      "grad_norm": 31.875,
+      "grad_norm_var": 7.4228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3327,
+      "loss/crossentropy": 2.102003552019596,
+      "loss/hidden": 3.386328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19359097983688117,
+      "step": 20460
+    },
+    {
+      "epoch": 0.51175,
+      "grad_norm": 33.0,
+      "grad_norm_var": 5.968684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3575,
+      "loss/crossentropy": 2.117770975828171,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20843392983078957,
+      "step": 20470
+    },
+    {
+      "epoch": 0.512,
+      "grad_norm": 28.625,
+      "grad_norm_var": 3.80390625,
+      "learning_rate": 0.0001,
+      "loss": 7.294,
+      "loss/crossentropy": 2.048271709680557,
+      "loss/hidden": 3.3421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19015705585479736,
+      "step": 20480
+    },
+    {
+      "epoch": 0.51225,
+      "grad_norm": 27.625,
+      "grad_norm_var": 4.162955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2999,
+      "loss/crossentropy": 2.2269428968429565,
+      "loss/hidden": 3.34921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18866086583584546,
+      "step": 20490
+    },
+    {
+      "epoch": 0.5125,
+      "grad_norm": 30.5,
+      "grad_norm_var": 6.625455729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2051,
+      "loss/crossentropy": 2.077080491185188,
+      "loss/hidden": 3.29921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18723124526441098,
+      "step": 20500
+    },
+    {
+      "epoch": 0.51275,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.943489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3489,
+      "loss/crossentropy": 1.9472593426704408,
+      "loss/hidden": 3.421484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18634757827967405,
+      "step": 20510
+    },
+    {
+      "epoch": 0.513,
+      "grad_norm": 29.875,
+      "grad_norm_var": 4.5431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2731,
+      "loss/crossentropy": 2.0889667838811876,
+      "loss/hidden": 3.283203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17469800189137458,
+      "step": 20520
+    },
+    {
+      "epoch": 0.51325,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.168684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3807,
+      "loss/crossentropy": 2.1357873290777207,
+      "loss/hidden": 3.406640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19474379923194646,
+      "step": 20530
+    },
+    {
+      "epoch": 0.5135,
+      "grad_norm": 28.625,
+      "grad_norm_var": 3.7177083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3564,
+      "loss/crossentropy": 2.1814554005861284,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19765354953706266,
+      "step": 20540
+    },
+    {
+      "epoch": 0.51375,
+      "grad_norm": 29.5,
+      "grad_norm_var": 7.0916015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3209,
+      "loss/crossentropy": 1.9481236249208451,
+      "loss/hidden": 3.366796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17467648405581712,
+      "step": 20550
+    },
+    {
+      "epoch": 0.514,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.842643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2827,
+      "loss/crossentropy": 1.9511701211333274,
+      "loss/hidden": 3.479296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17900633215904235,
+      "step": 20560
+    },
+    {
+      "epoch": 0.51425,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.1205729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3626,
+      "loss/crossentropy": 1.947856567800045,
+      "loss/hidden": 3.38984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1868027526885271,
+      "step": 20570
+    },
+    {
+      "epoch": 0.5145,
+      "grad_norm": 28.625,
+      "grad_norm_var": 19.1875,
+      "learning_rate": 0.0001,
+      "loss": 7.2629,
+      "loss/crossentropy": 2.0266165107488634,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17448003627359868,
+      "step": 20580
+    },
+    {
+      "epoch": 0.51475,
+      "grad_norm": 30.0,
+      "grad_norm_var": 0.7046223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2807,
+      "loss/crossentropy": 2.015270346403122,
+      "loss/hidden": 3.383203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18326627276837826,
+      "step": 20590
+    },
+    {
+      "epoch": 0.515,
+      "grad_norm": 31.875,
+      "grad_norm_var": 5.234309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3668,
+      "loss/crossentropy": 2.034985066950321,
+      "loss/hidden": 3.3265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17759786024689675,
+      "step": 20600
+    },
+    {
+      "epoch": 0.51525,
+      "grad_norm": 30.5,
+      "grad_norm_var": 6.9494140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3519,
+      "loss/crossentropy": 2.0228724449872972,
+      "loss/hidden": 3.31328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18106682635843754,
+      "step": 20610
+    },
+    {
+      "epoch": 0.5155,
+      "grad_norm": 48.0,
+      "grad_norm_var": 41.123958333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4442,
+      "loss/crossentropy": 1.99410739839077,
+      "loss/hidden": 3.528515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18867430798709392,
+      "step": 20620
+    },
+    {
+      "epoch": 0.51575,
+      "grad_norm": 31.0,
+      "grad_norm_var": 112.25807291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3746,
+      "loss/crossentropy": 2.063592994213104,
+      "loss/hidden": 3.43671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20251007843762636,
+      "step": 20630
+    },
+    {
+      "epoch": 0.516,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.1122395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3186,
+      "loss/crossentropy": 2.1162465453147887,
+      "loss/hidden": 3.2609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1863781152293086,
+      "step": 20640
+    },
+    {
+      "epoch": 0.51625,
+      "grad_norm": 29.375,
+      "grad_norm_var": 11.701041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.303,
+      "loss/crossentropy": 1.9705959290266037,
+      "loss/hidden": 3.4875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.189107047021389,
+      "step": 20650
+    },
+    {
+      "epoch": 0.5165,
+      "grad_norm": 42.5,
+      "grad_norm_var": 25.134830729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2953,
+      "loss/crossentropy": 2.04353059977293,
+      "loss/hidden": 3.3453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18167578242719173,
+      "step": 20660
+    },
+    {
+      "epoch": 0.51675,
+      "grad_norm": 63.25,
+      "grad_norm_var": 90.03483072916667,
+      "learning_rate": 0.0001,
+      "loss": 7.1868,
+      "loss/crossentropy": 1.9600037515163422,
+      "loss/hidden": 3.330859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1769449569284916,
+      "step": 20670
+    },
+    {
+      "epoch": 0.517,
+      "grad_norm": 28.75,
+      "grad_norm_var": 73.69264322916666,
+      "learning_rate": 0.0001,
+      "loss": 7.3364,
+      "loss/crossentropy": 2.235697329044342,
+      "loss/hidden": 3.41796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19675933755934238,
+      "step": 20680
+    },
+    {
+      "epoch": 0.51725,
+      "grad_norm": 60.0,
+      "grad_norm_var": 67.38951822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.2911,
+      "loss/crossentropy": 2.047742946445942,
+      "loss/hidden": 3.398828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1932275788858533,
+      "step": 20690
+    },
+    {
+      "epoch": 0.5175,
+      "grad_norm": 30.875,
+      "grad_norm_var": 66.7978515625,
+      "learning_rate": 0.0001,
+      "loss": 7.369,
+      "loss/crossentropy": 2.1426502510905268,
+      "loss/hidden": 3.38671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19659061841666697,
+      "step": 20700
+    },
+    {
+      "epoch": 0.51775,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.9848307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3526,
+      "loss/crossentropy": 2.0521486185491087,
+      "loss/hidden": 3.31484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18707139398902656,
+      "step": 20710
+    },
+    {
+      "epoch": 0.518,
+      "grad_norm": 30.5,
+      "grad_norm_var": 16.7634765625,
+      "learning_rate": 0.0001,
+      "loss": 7.363,
+      "loss/crossentropy": 2.0975049562752246,
+      "loss/hidden": 3.338671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.190740005671978,
+      "step": 20720
+    },
+    {
+      "epoch": 0.51825,
+      "grad_norm": 34.5,
+      "grad_norm_var": 27.069205729166665,
+      "learning_rate": 0.0001,
+      "loss": 7.3643,
+      "loss/crossentropy": 2.1847240805625914,
+      "loss/hidden": 3.42265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20393175166100264,
+      "step": 20730
+    },
+    {
+      "epoch": 0.5185,
+      "grad_norm": 28.5,
+      "grad_norm_var": 28.865625,
+      "learning_rate": 0.0001,
+      "loss": 7.3442,
+      "loss/crossentropy": 2.079416597634554,
+      "loss/hidden": 3.362890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18700419086962938,
+      "step": 20740
+    },
+    {
+      "epoch": 0.51875,
+      "grad_norm": 30.0,
+      "grad_norm_var": 5.898893229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2968,
+      "loss/crossentropy": 2.099186307191849,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19339050203561783,
+      "step": 20750
+    },
+    {
+      "epoch": 0.519,
+      "grad_norm": 33.5,
+      "grad_norm_var": 7.76640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2292,
+      "loss/crossentropy": 1.9471840865910053,
+      "loss/hidden": 3.357421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18490714272484182,
+      "step": 20760
+    },
+    {
+      "epoch": 0.51925,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.6176432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2505,
+      "loss/crossentropy": 1.9791502140462398,
+      "loss/hidden": 3.315625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1867075043730438,
+      "step": 20770
+    },
+    {
+      "epoch": 0.5195,
+      "grad_norm": 28.25,
+      "grad_norm_var": 1.8843098958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2881,
+      "loss/crossentropy": 2.228019216656685,
+      "loss/hidden": 3.33828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19135302156209946,
+      "step": 20780
+    },
+    {
+      "epoch": 0.51975,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.9684895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2552,
+      "loss/crossentropy": 2.0987790375947952,
+      "loss/hidden": 3.3421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17536395322531462,
+      "step": 20790
+    },
+    {
+      "epoch": 0.52,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.2309895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3921,
+      "loss/crossentropy": 2.0298918724060058,
+      "loss/hidden": 3.38359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1898334242403507,
+      "step": 20800
+    },
+    {
+      "epoch": 0.52025,
+      "grad_norm": 28.875,
+      "grad_norm_var": 1.5083333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.234,
+      "loss/crossentropy": 1.9918959826231002,
+      "loss/hidden": 3.3140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18577282279729843,
+      "step": 20810
+    },
+    {
+      "epoch": 0.5205,
+      "grad_norm": 34.25,
+      "grad_norm_var": 3.3785807291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3487,
+      "loss/crossentropy": 2.0936102479696275,
+      "loss/hidden": 3.356640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19174492321908473,
+      "step": 20820
+    },
+    {
+      "epoch": 0.52075,
+      "grad_norm": 37.75,
+      "grad_norm_var": 1.9157890389214495e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2159,
+      "loss/crossentropy": 1.8458645723760128,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1738974804058671,
+      "step": 20830
+    },
+    {
+      "epoch": 0.521,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.9157890383966372e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2727,
+      "loss/crossentropy": 1.9869740903377533,
+      "loss/hidden": 3.324609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19819957707077265,
+      "step": 20840
+    },
+    {
+      "epoch": 0.52125,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.275,
+      "learning_rate": 0.0001,
+      "loss": 7.2681,
+      "loss/crossentropy": 2.0040796995162964,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18156437613070012,
+      "step": 20850
+    },
+    {
+      "epoch": 0.5215,
+      "grad_norm": 29.375,
+      "grad_norm_var": 4.849934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3494,
+      "loss/crossentropy": 2.165298455953598,
+      "loss/hidden": 3.26796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18115576189011334,
+      "step": 20860
+    },
+    {
+      "epoch": 0.52175,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.8848307291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3202,
+      "loss/crossentropy": 2.100299298763275,
+      "loss/hidden": 3.37421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19552922286093236,
+      "step": 20870
+    },
+    {
+      "epoch": 0.522,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.60390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2684,
+      "loss/crossentropy": 2.022940080612898,
+      "loss/hidden": 3.255859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1667741946876049,
+      "step": 20880
+    },
+    {
+      "epoch": 0.52225,
+      "grad_norm": 55.25,
+      "grad_norm_var": 40.665625,
+      "learning_rate": 0.0001,
+      "loss": 7.2855,
+      "loss/crossentropy": 2.102290881425142,
+      "loss/hidden": 3.313671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1791887005791068,
+      "step": 20890
+    },
+    {
+      "epoch": 0.5225,
+      "grad_norm": 30.375,
+      "grad_norm_var": 44.436458333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1947,
+      "loss/crossentropy": 1.8682746514678001,
+      "loss/hidden": 3.37109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18195230644196272,
+      "step": 20900
+    },
+    {
+      "epoch": 0.52275,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.5416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3462,
+      "loss/crossentropy": 2.0893214523792265,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19191989079117774,
+      "step": 20910
+    },
+    {
+      "epoch": 0.523,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.3520182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.347,
+      "loss/crossentropy": 2.0877039656043053,
+      "loss/hidden": 3.3640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1851141469553113,
+      "step": 20920
+    },
+    {
+      "epoch": 0.52325,
+      "grad_norm": 28.0,
+      "grad_norm_var": 2.8997395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2357,
+      "loss/crossentropy": 2.1176202185451984,
+      "loss/hidden": 3.404296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19145621228963136,
+      "step": 20930
+    },
+    {
+      "epoch": 0.5235,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.2186848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2674,
+      "loss/crossentropy": 2.0402719154953957,
+      "loss/hidden": 3.371875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1739909777417779,
+      "step": 20940
+    },
+    {
+      "epoch": 0.52375,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.0759765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3178,
+      "loss/crossentropy": 2.030729368329048,
+      "loss/hidden": 3.30859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1842687962576747,
+      "step": 20950
+    },
+    {
+      "epoch": 0.524,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.3958333333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3384,
+      "loss/crossentropy": 2.0869288608431815,
+      "loss/hidden": 3.352734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18400898948311806,
+      "step": 20960
+    },
+    {
+      "epoch": 0.52425,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.9067057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4186,
+      "loss/crossentropy": 2.11807075291872,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18031747452914715,
+      "step": 20970
+    },
+    {
+      "epoch": 0.5245,
+      "grad_norm": 29.75,
+      "grad_norm_var": 6.5244140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2164,
+      "loss/crossentropy": 1.9718056872487069,
+      "loss/hidden": 3.26640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17341692205518483,
+      "step": 20980
+    },
+    {
+      "epoch": 0.52475,
+      "grad_norm": 30.625,
+      "grad_norm_var": 6.382747395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3702,
+      "loss/crossentropy": 2.0978681847453116,
+      "loss/hidden": 3.321484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17709601484239101,
+      "step": 20990
+    },
+    {
+      "epoch": 0.525,
+      "grad_norm": 31.5,
+      "grad_norm_var": 9.713997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3043,
+      "loss/crossentropy": 2.034855252504349,
+      "loss/hidden": 3.491796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1906108744442463,
+      "step": 21000
+    },
+    {
+      "epoch": 0.52525,
+      "grad_norm": 28.625,
+      "grad_norm_var": 13.075455729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4447,
+      "loss/crossentropy": 2.2921934738755225,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1889318484812975,
+      "step": 21010
+    },
+    {
+      "epoch": 0.5255,
+      "grad_norm": 29.5,
+      "grad_norm_var": 6.339322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2976,
+      "loss/crossentropy": 1.8201854094862937,
+      "loss/hidden": 3.4828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18457184694707393,
+      "step": 21020
+    },
+    {
+      "epoch": 0.52575,
+      "grad_norm": 28.875,
+      "grad_norm_var": 29.380989583333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3322,
+      "loss/crossentropy": 2.131899508088827,
+      "loss/hidden": 3.43515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2020415196195245,
+      "step": 21030
+    },
+    {
+      "epoch": 0.526,
+      "grad_norm": 31.875,
+      "grad_norm_var": 41.7431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2982,
+      "loss/crossentropy": 1.802038711309433,
+      "loss/hidden": 3.43203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1719184698536992,
+      "step": 21040
+    },
+    {
+      "epoch": 0.52625,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.9468098958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2365,
+      "loss/crossentropy": 1.9720843493938447,
+      "loss/hidden": 3.38046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18162036091089248,
+      "step": 21050
+    },
+    {
+      "epoch": 0.5265,
+      "grad_norm": 29.625,
+      "grad_norm_var": 8.392708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2629,
+      "loss/crossentropy": 2.101500564068556,
+      "loss/hidden": 3.3046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18104594042524697,
+      "step": 21060
+    },
+    {
+      "epoch": 0.52675,
+      "grad_norm": 31.75,
+      "grad_norm_var": 12.012239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4139,
+      "loss/crossentropy": 2.192843732237816,
+      "loss/hidden": 3.43671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1871086286380887,
+      "step": 21070
+    },
+    {
+      "epoch": 0.527,
+      "grad_norm": 33.75,
+      "grad_norm_var": 7.6291015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3464,
+      "loss/crossentropy": 1.935421773791313,
+      "loss/hidden": 3.371484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18491943813860418,
+      "step": 21080
+    },
+    {
+      "epoch": 0.52725,
+      "grad_norm": 29.25,
+      "grad_norm_var": 8.599934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3769,
+      "loss/crossentropy": 2.1220942765474318,
+      "loss/hidden": 3.29609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18080089893192053,
+      "step": 21090
+    },
+    {
+      "epoch": 0.5275,
+      "grad_norm": 32.5,
+      "grad_norm_var": 13.941666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3172,
+      "loss/crossentropy": 2.1950094163417817,
+      "loss/hidden": 3.432421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19560035392642022,
+      "step": 21100
+    },
+    {
+      "epoch": 0.52775,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.9431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2835,
+      "loss/crossentropy": 2.0231982678174973,
+      "loss/hidden": 3.41015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1959500327706337,
+      "step": 21110
+    },
+    {
+      "epoch": 0.528,
+      "grad_norm": 34.75,
+      "grad_norm_var": 4.874934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2968,
+      "loss/crossentropy": 2.014877498149872,
+      "loss/hidden": 3.38828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18462858479470015,
+      "step": 21120
+    },
+    {
+      "epoch": 0.52825,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.410872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2742,
+      "loss/crossentropy": 2.0856463953852655,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1750152625143528,
+      "step": 21130
+    },
+    {
+      "epoch": 0.5285,
+      "grad_norm": 30.5,
+      "grad_norm_var": 4.5125,
+      "learning_rate": 0.0001,
+      "loss": 7.3537,
+      "loss/crossentropy": 2.097464480996132,
+      "loss/hidden": 3.309375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1879219425842166,
+      "step": 21140
+    },
+    {
+      "epoch": 0.52875,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.91875,
+      "learning_rate": 0.0001,
+      "loss": 7.2096,
+      "loss/crossentropy": 2.025162447988987,
+      "loss/hidden": 3.305859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18031715657562017,
+      "step": 21150
+    },
+    {
+      "epoch": 0.529,
+      "grad_norm": 31.0,
+      "grad_norm_var": 687.7858723958333,
+      "learning_rate": 0.0001,
+      "loss": 7.3077,
+      "loss/crossentropy": 2.0105077303946017,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17551136016845703,
+      "step": 21160
+    },
+    {
+      "epoch": 0.52925,
+      "grad_norm": 27.875,
+      "grad_norm_var": 690.9580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2711,
+      "loss/crossentropy": 1.984055256843567,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18593314290046692,
+      "step": 21170
+    },
+    {
+      "epoch": 0.5295,
+      "grad_norm": 30.125,
+      "grad_norm_var": 6.970768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2034,
+      "loss/crossentropy": 2.1591003805398943,
+      "loss/hidden": 3.317578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18686197344213723,
+      "step": 21180
+    },
+    {
+      "epoch": 0.52975,
+      "grad_norm": 27.625,
+      "grad_norm_var": 4.736393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4247,
+      "loss/crossentropy": 2.0217775389552117,
+      "loss/hidden": 3.347265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1973855821415782,
+      "step": 21190
+    },
+    {
+      "epoch": 0.53,
+      "grad_norm": 28.75,
+      "grad_norm_var": 5.764322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3758,
+      "loss/crossentropy": 2.0906185880303383,
+      "loss/hidden": 3.4078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19189006537199021,
+      "step": 21200
+    },
+    {
+      "epoch": 0.53025,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.7018229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3831,
+      "loss/crossentropy": 2.2920484393835068,
+      "loss/hidden": 3.34296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19543348867446184,
+      "step": 21210
+    },
+    {
+      "epoch": 0.5305,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.301540273547833e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3275,
+      "loss/crossentropy": 2.0418326549232004,
+      "loss/hidden": 3.40546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1798579154536128,
+      "step": 21220
+    },
+    {
+      "epoch": 0.53075,
+      "grad_norm": 30.625,
+      "grad_norm_var": 17.545572916666668,
+      "learning_rate": 0.0001,
+      "loss": 7.3427,
+      "loss/crossentropy": 2.1924046859145165,
+      "loss/hidden": 3.330859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1793345332145691,
+      "step": 21230
+    },
+    {
+      "epoch": 0.531,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.9702473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1754,
+      "loss/crossentropy": 1.9836316756904124,
+      "loss/hidden": 3.3453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18065028581768275,
+      "step": 21240
+    },
+    {
+      "epoch": 0.53125,
+      "grad_norm": 28.625,
+      "grad_norm_var": 1.4410807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3525,
+      "loss/crossentropy": 1.9980021633207798,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18613223265856504,
+      "step": 21250
+    },
+    {
+      "epoch": 0.5315,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.5712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3235,
+      "loss/crossentropy": 1.9790296763181687,
+      "loss/hidden": 3.471875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19255476742982863,
+      "step": 21260
+    },
+    {
+      "epoch": 0.53175,
+      "grad_norm": 28.375,
+      "grad_norm_var": 3.1264973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2533,
+      "loss/crossentropy": 2.0499105736613275,
+      "loss/hidden": 3.32890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17501375842839478,
+      "step": 21270
+    },
+    {
+      "epoch": 0.532,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.6997395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2437,
+      "loss/crossentropy": 2.0615583658218384,
+      "loss/hidden": 3.273828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17125342395156623,
+      "step": 21280
+    },
+    {
+      "epoch": 0.53225,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.3228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2509,
+      "loss/crossentropy": 2.0179983586072923,
+      "loss/hidden": 3.30546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17567140907049178,
+      "step": 21290
+    },
+    {
+      "epoch": 0.5325,
+      "grad_norm": 28.875,
+      "grad_norm_var": 0.8462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2822,
+      "loss/crossentropy": 2.044732092320919,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17829035893082618,
+      "step": 21300
+    },
+    {
+      "epoch": 0.53275,
+      "grad_norm": 30.625,
+      "grad_norm_var": 6.9916015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2366,
+      "loss/crossentropy": 2.108456004410982,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1892499382607639,
+      "step": 21310
+    },
+    {
+      "epoch": 0.533,
+      "grad_norm": 30.75,
+      "grad_norm_var": 7.531184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2825,
+      "loss/crossentropy": 2.0786417722702026,
+      "loss/hidden": 3.3546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1849788771942258,
+      "step": 21320
+    },
+    {
+      "epoch": 0.53325,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.8067057291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3622,
+      "loss/crossentropy": 2.0772973991930486,
+      "loss/hidden": 3.391015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19109403509646655,
+      "step": 21330
+    },
+    {
+      "epoch": 0.5335,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.414322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3283,
+      "loss/crossentropy": 2.2267979592084886,
+      "loss/hidden": 3.263671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17751519959419965,
+      "step": 21340
+    },
+    {
+      "epoch": 0.53375,
+      "grad_norm": 31.0,
+      "grad_norm_var": 7.658333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3284,
+      "loss/crossentropy": 2.0170826002955438,
+      "loss/hidden": 3.347265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17177296355366706,
+      "step": 21350
+    },
+    {
+      "epoch": 0.534,
+      "grad_norm": 29.0,
+      "grad_norm_var": 1.7,
+      "learning_rate": 0.0001,
+      "loss": 7.267,
+      "loss/crossentropy": 2.175469179451466,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2005373438820243,
+      "step": 21360
+    },
+    {
+      "epoch": 0.53425,
+      "grad_norm": 30.625,
+      "grad_norm_var": 8.731705729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.137,
+      "loss/crossentropy": 2.0817506462335587,
+      "loss/hidden": 3.30625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17554874606430532,
+      "step": 21370
+    },
+    {
+      "epoch": 0.5345,
+      "grad_norm": 29.375,
+      "grad_norm_var": 10.517122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3312,
+      "loss/crossentropy": 2.0762263566255568,
+      "loss/hidden": 3.244140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17292992006987334,
+      "step": 21380
+    },
+    {
+      "epoch": 0.53475,
+      "grad_norm": 29.375,
+      "grad_norm_var": 7.6447265625,
+      "learning_rate": 0.0001,
+      "loss": 7.279,
+      "loss/crossentropy": 2.073785340040922,
+      "loss/hidden": 3.2546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18681957013905048,
+      "step": 21390
+    },
+    {
+      "epoch": 0.535,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.1333333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3367,
+      "loss/crossentropy": 2.16663718521595,
+      "loss/hidden": 3.296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18254186888225377,
+      "step": 21400
+    },
+    {
+      "epoch": 0.53525,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.27890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3824,
+      "loss/crossentropy": 2.000949743390083,
+      "loss/hidden": 3.419140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19326272774487735,
+      "step": 21410
+    },
+    {
+      "epoch": 0.5355,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.3202473958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3294,
+      "loss/crossentropy": 1.9899908214807511,
+      "loss/hidden": 3.369140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19131155535578728,
+      "step": 21420
+    },
+    {
+      "epoch": 0.53575,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.3811848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3148,
+      "loss/crossentropy": 2.1891008853912353,
+      "loss/hidden": 3.40078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19048418197780848,
+      "step": 21430
+    },
+    {
+      "epoch": 0.536,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.6072916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2268,
+      "loss/crossentropy": 2.1451114103198052,
+      "loss/hidden": 3.33203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.179903282225132,
+      "step": 21440
+    },
+    {
+      "epoch": 0.53625,
+      "grad_norm": 31.125,
+      "grad_norm_var": 5.038997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3707,
+      "loss/crossentropy": 2.0621043920516966,
+      "loss/hidden": 3.501953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2030002610757947,
+      "step": 21450
+    },
+    {
+      "epoch": 0.5365,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.1434895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3413,
+      "loss/crossentropy": 2.0637945592403413,
+      "loss/hidden": 3.440625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21434453465044498,
+      "step": 21460
+    },
+    {
+      "epoch": 0.53675,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.9046223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.331,
+      "loss/crossentropy": 2.100746136903763,
+      "loss/hidden": 3.4203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20721383020281792,
+      "step": 21470
+    },
+    {
+      "epoch": 0.537,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.8875,
+      "learning_rate": 0.0001,
+      "loss": 7.266,
+      "loss/crossentropy": 2.085133180767298,
+      "loss/hidden": 3.412890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19881708025932313,
+      "step": 21480
+    },
+    {
+      "epoch": 0.53725,
+      "grad_norm": 26.875,
+      "grad_norm_var": 4.8931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2806,
+      "loss/crossentropy": 1.9575376056134701,
+      "loss/hidden": 3.312890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17017574273049832,
+      "step": 21490
+    },
+    {
+      "epoch": 0.5375,
+      "grad_norm": 36.75,
+      "grad_norm_var": 27.210872395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3797,
+      "loss/crossentropy": 2.107981327176094,
+      "loss/hidden": 3.496484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20202290676534176,
+      "step": 21500
+    },
+    {
+      "epoch": 0.53775,
+      "grad_norm": 43.25,
+      "grad_norm_var": 217.90625,
+      "learning_rate": 0.0001,
+      "loss": 7.3549,
+      "loss/crossentropy": 2.0315024718642234,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18244437091052532,
+      "step": 21510
+    },
+    {
+      "epoch": 0.538,
+      "grad_norm": 35.75,
+      "grad_norm_var": 209.64140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2932,
+      "loss/crossentropy": 2.0473300129175187,
+      "loss/hidden": 3.336328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1864867027848959,
+      "step": 21520
+    },
+    {
+      "epoch": 0.53825,
+      "grad_norm": 30.0,
+      "grad_norm_var": 28.510872395833335,
+      "learning_rate": 0.0001,
+      "loss": 7.2708,
+      "loss/crossentropy": 2.2367413520812987,
+      "loss/hidden": 3.365625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19972213320434093,
+      "step": 21530
+    },
+    {
+      "epoch": 0.5385,
+      "grad_norm": 30.875,
+      "grad_norm_var": 30.212955729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2704,
+      "loss/crossentropy": 2.1623730801045893,
+      "loss/hidden": 3.321875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1971738815307617,
+      "step": 21540
+    },
+    {
+      "epoch": 0.53875,
+      "grad_norm": 35.0,
+      "grad_norm_var": 7.9134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2673,
+      "loss/crossentropy": 2.1150405749678614,
+      "loss/hidden": 3.287109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18175593428313733,
+      "step": 21550
+    },
+    {
+      "epoch": 0.539,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.928465630196033e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3405,
+      "loss/crossentropy": 2.0370446920394896,
+      "loss/hidden": 3.30703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1800082778558135,
+      "step": 21560
+    },
+    {
+      "epoch": 0.53925,
+      "grad_norm": 35.5,
+      "grad_norm_var": 2.928465629582826e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3808,
+      "loss/crossentropy": 2.1404812157154085,
+      "loss/hidden": 3.35703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18585823830217124,
+      "step": 21570
+    },
+    {
+      "epoch": 0.5395,
+      "grad_norm": 30.375,
+      "grad_norm_var": 4.48515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3203,
+      "loss/crossentropy": 2.2115766972303392,
+      "loss/hidden": 3.409765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19159688670188188,
+      "step": 21580
+    },
+    {
+      "epoch": 0.53975,
+      "grad_norm": 28.0,
+      "grad_norm_var": 6.1666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2824,
+      "loss/crossentropy": 2.1206958115100862,
+      "loss/hidden": 3.3875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18236497584730388,
+      "step": 21590
+    },
+    {
+      "epoch": 0.54,
+      "grad_norm": 33.0,
+      "grad_norm_var": 8.221809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1904,
+      "loss/crossentropy": 1.9439353846013545,
+      "loss/hidden": 3.3125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16596114486455918,
+      "step": 21600
+    },
+    {
+      "epoch": 0.54025,
+      "grad_norm": 30.625,
+      "grad_norm_var": 4.26640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2192,
+      "loss/crossentropy": 2.084056280553341,
+      "loss/hidden": 3.39921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1889295268803835,
+      "step": 21610
+    },
+    {
+      "epoch": 0.5405,
+      "grad_norm": 30.75,
+      "grad_norm_var": 4.08515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2147,
+      "loss/crossentropy": 1.9731713935732842,
+      "loss/hidden": 3.276953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17223210986703635,
+      "step": 21620
+    },
+    {
+      "epoch": 0.54075,
+      "grad_norm": 30.875,
+      "grad_norm_var": 8.837239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2933,
+      "loss/crossentropy": 2.049588477611542,
+      "loss/hidden": 3.321875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18233908694237472,
+      "step": 21630
+    },
+    {
+      "epoch": 0.541,
+      "grad_norm": 29.625,
+      "grad_norm_var": 5.414322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2083,
+      "loss/crossentropy": 2.045769859850407,
+      "loss/hidden": 3.2625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17521222569048406,
+      "step": 21640
+    },
+    {
+      "epoch": 0.54125,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.04140625,
+      "learning_rate": 0.0001,
+      "loss": 7.297,
+      "loss/crossentropy": 2.1753006681799887,
+      "loss/hidden": 3.353515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1932337900623679,
+      "step": 21650
+    },
+    {
+      "epoch": 0.5415,
+      "grad_norm": 27.875,
+      "grad_norm_var": 4.039583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1617,
+      "loss/crossentropy": 2.1425692200660706,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18889971654862164,
+      "step": 21660
+    },
+    {
+      "epoch": 0.54175,
+      "grad_norm": 34.25,
+      "grad_norm_var": 4.156184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1944,
+      "loss/crossentropy": 2.0136908262968065,
+      "loss/hidden": 3.2953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1680521246045828,
+      "step": 21670
+    },
+    {
+      "epoch": 0.542,
+      "grad_norm": 30.875,
+      "grad_norm_var": 4.99140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2897,
+      "loss/crossentropy": 2.148314264416695,
+      "loss/hidden": 3.24765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.170760527998209,
+      "step": 21680
+    },
+    {
+      "epoch": 0.54225,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.8059895833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.1925,
+      "loss/crossentropy": 2.055006366968155,
+      "loss/hidden": 3.375390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1854418007656932,
+      "step": 21690
+    },
+    {
+      "epoch": 0.5425,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.6384765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3447,
+      "loss/crossentropy": 2.0953226678073404,
+      "loss/hidden": 3.386328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1921064408496022,
+      "step": 21700
+    },
+    {
+      "epoch": 0.54275,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.0625,
+      "learning_rate": 0.0001,
+      "loss": 7.3351,
+      "loss/crossentropy": 2.075012197345495,
+      "loss/hidden": 3.371484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18619187362492085,
+      "step": 21710
+    },
+    {
+      "epoch": 0.543,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.9306640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2675,
+      "loss/crossentropy": 2.104280537366867,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18965906854718922,
+      "step": 21720
+    },
+    {
+      "epoch": 0.54325,
+      "grad_norm": 28.625,
+      "grad_norm_var": 4.276041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2077,
+      "loss/crossentropy": 1.8971405908465386,
+      "loss/hidden": 3.35390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17859434587880968,
+      "step": 21730
+    },
+    {
+      "epoch": 0.5435,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.109375,
+      "learning_rate": 0.0001,
+      "loss": 7.3256,
+      "loss/crossentropy": 2.1083227381110192,
+      "loss/hidden": 3.386328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1904505856335163,
+      "step": 21740
+    },
+    {
+      "epoch": 0.54375,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.4728515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3875,
+      "loss/crossentropy": 2.3079893991351126,
+      "loss/hidden": 3.28046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18368349261581898,
+      "step": 21750
+    },
+    {
+      "epoch": 0.544,
+      "grad_norm": 30.125,
+      "grad_norm_var": 5.176822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3763,
+      "loss/crossentropy": 2.098197969794273,
+      "loss/hidden": 3.392578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2088245895691216,
+      "step": 21760
+    },
+    {
+      "epoch": 0.54425,
+      "grad_norm": 29.5,
+      "grad_norm_var": 7.164322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3008,
+      "loss/crossentropy": 2.1331639789044856,
+      "loss/hidden": 3.353125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17733938563615084,
+      "step": 21770
+    },
+    {
+      "epoch": 0.5445,
+      "grad_norm": 30.625,
+      "grad_norm_var": 165.75045572916667,
+      "learning_rate": 0.0001,
+      "loss": 7.2084,
+      "loss/crossentropy": 1.9805914640426636,
+      "loss/hidden": 3.355859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18551983144134282,
+      "step": 21780
+    },
+    {
+      "epoch": 0.54475,
+      "grad_norm": 29.125,
+      "grad_norm_var": 32.90598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2898,
+      "loss/crossentropy": 2.0330356270074845,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18011851627379655,
+      "step": 21790
+    },
+    {
+      "epoch": 0.545,
+      "grad_norm": 37.25,
+      "grad_norm_var": 29.716666666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4356,
+      "loss/crossentropy": 2.1713989078998566,
+      "loss/hidden": 3.334375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19549211151897908,
+      "step": 21800
+    },
+    {
+      "epoch": 0.54525,
+      "grad_norm": 29.75,
+      "grad_norm_var": 6.017122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2726,
+      "loss/crossentropy": 2.2081006675958634,
+      "loss/hidden": 3.298828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1883497454226017,
+      "step": 21810
+    },
+    {
+      "epoch": 0.5455,
+      "grad_norm": 31.25,
+      "grad_norm_var": 4.859830729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3724,
+      "loss/crossentropy": 1.9547579608857633,
+      "loss/hidden": 3.32109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1771907014772296,
+      "step": 21820
+    },
+    {
+      "epoch": 0.54575,
+      "grad_norm": 32.75,
+      "grad_norm_var": 5.364518229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3612,
+      "loss/crossentropy": 2.098206965625286,
+      "loss/hidden": 3.459765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19235289450734855,
+      "step": 21830
+    },
+    {
+      "epoch": 0.546,
+      "grad_norm": 31.5,
+      "grad_norm_var": 6.561393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2737,
+      "loss/crossentropy": 2.1263977140188217,
+      "loss/hidden": 3.41796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1888118337839842,
+      "step": 21840
+    },
+    {
+      "epoch": 0.54625,
+      "grad_norm": 33.0,
+      "grad_norm_var": 5.3603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2304,
+      "loss/crossentropy": 2.101663938164711,
+      "loss/hidden": 3.27265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17398083284497262,
+      "step": 21850
+    },
+    {
+      "epoch": 0.5465,
+      "grad_norm": 28.375,
+      "grad_norm_var": 4.398958333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1952,
+      "loss/crossentropy": 1.9339447408914565,
+      "loss/hidden": 3.331640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16659772507846354,
+      "step": 21860
+    },
+    {
+      "epoch": 0.54675,
+      "grad_norm": 35.5,
+      "grad_norm_var": 4.6666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2218,
+      "loss/crossentropy": 1.9582382440567017,
+      "loss/hidden": 3.279296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17305021807551385,
+      "step": 21870
+    },
+    {
+      "epoch": 0.547,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.503059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1931,
+      "loss/crossentropy": 1.9935566574335097,
+      "loss/hidden": 3.27734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17071465943008662,
+      "step": 21880
+    },
+    {
+      "epoch": 0.54725,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.862955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2374,
+      "loss/crossentropy": 2.036357142031193,
+      "loss/hidden": 3.424609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19538825107738375,
+      "step": 21890
+    },
+    {
+      "epoch": 0.5475,
+      "grad_norm": 32.25,
+      "grad_norm_var": 5.046875,
+      "learning_rate": 0.0001,
+      "loss": 7.2621,
+      "loss/crossentropy": 1.92364434376359,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17744148215278982,
+      "step": 21900
+    },
+    {
+      "epoch": 0.54775,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.564583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.285,
+      "loss/crossentropy": 2.0298509031534193,
+      "loss/hidden": 3.234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17162161748856306,
+      "step": 21910
+    },
+    {
+      "epoch": 0.548,
+      "grad_norm": 31.0,
+      "grad_norm_var": 5.1619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3575,
+      "loss/crossentropy": 2.175525971502066,
+      "loss/hidden": 3.298046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18408444514498115,
+      "step": 21920
+    },
+    {
+      "epoch": 0.54825,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.9114583333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.1831,
+      "loss/crossentropy": 1.8858811140060425,
+      "loss/hidden": 3.418359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1825306786224246,
+      "step": 21930
+    },
+    {
+      "epoch": 0.5485,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.62890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2268,
+      "loss/crossentropy": 2.026767262816429,
+      "loss/hidden": 3.30234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17721015885472297,
+      "step": 21940
+    },
+    {
+      "epoch": 0.54875,
+      "grad_norm": 36.5,
+      "grad_norm_var": 5.305989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2989,
+      "loss/crossentropy": 1.9967989712953567,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1873170206323266,
+      "step": 21950
+    },
+    {
+      "epoch": 0.549,
+      "grad_norm": 28.375,
+      "grad_norm_var": 239.72180989583333,
+      "learning_rate": 0.0001,
+      "loss": 7.3301,
+      "loss/crossentropy": 2.163920529931784,
+      "loss/hidden": 3.39921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1893145816400647,
+      "step": 21960
+    },
+    {
+      "epoch": 0.54925,
+      "grad_norm": 29.0,
+      "grad_norm_var": 4.773372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3758,
+      "loss/crossentropy": 2.1935864582657816,
+      "loss/hidden": 3.358203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19310898408293725,
+      "step": 21970
+    },
+    {
+      "epoch": 0.5495,
+      "grad_norm": 36.0,
+      "grad_norm_var": 6.3072265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2928,
+      "loss/crossentropy": 1.9404637925326824,
+      "loss/hidden": 3.325,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18799830228090286,
+      "step": 21980
+    },
+    {
+      "epoch": 0.54975,
+      "grad_norm": 6174015488.0,
+      "grad_norm_var": 2.3824041793920087e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2622,
+      "loss/crossentropy": 2.111705342680216,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2000083826482296,
+      "step": 21990
+    },
+    {
+      "epoch": 0.55,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.3824041760992005e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3763,
+      "loss/crossentropy": 2.0126213818788528,
+      "loss/hidden": 3.427734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21517845019698142,
+      "step": 22000
+    },
+    {
+      "epoch": 0.55025,
+      "grad_norm": 29.625,
+      "grad_norm_var": 11.349934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3056,
+      "loss/crossentropy": 2.1753602206707,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1973609359934926,
+      "step": 22010
+    },
+    {
+      "epoch": 0.5505,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.3613932291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2592,
+      "loss/crossentropy": 2.045438584685326,
+      "loss/hidden": 3.287109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17302108742296696,
+      "step": 22020
+    },
+    {
+      "epoch": 0.55075,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.4893229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2572,
+      "loss/crossentropy": 1.988423915207386,
+      "loss/hidden": 3.345703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19198302906006576,
+      "step": 22030
+    },
+    {
+      "epoch": 0.551,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.018684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3145,
+      "loss/crossentropy": 2.186215503513813,
+      "loss/hidden": 3.321484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.188802495226264,
+      "step": 22040
+    },
+    {
+      "epoch": 0.55125,
+      "grad_norm": 35.5,
+      "grad_norm_var": 3.05,
+      "learning_rate": 0.0001,
+      "loss": 7.3018,
+      "loss/crossentropy": 2.0135563641786574,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19467689003795385,
+      "step": 22050
+    },
+    {
+      "epoch": 0.5515,
+      "grad_norm": 34.5,
+      "grad_norm_var": 34.940625,
+      "learning_rate": 0.0001,
+      "loss": 7.2785,
+      "loss/crossentropy": 1.9395711533725262,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1815367082133889,
+      "step": 22060
+    },
+    {
+      "epoch": 0.55175,
+      "grad_norm": 29.75,
+      "grad_norm_var": 23.817122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3435,
+      "loss/crossentropy": 2.024968159198761,
+      "loss/hidden": 3.51484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20050131883472205,
+      "step": 22070
+    },
+    {
+      "epoch": 0.552,
+      "grad_norm": 27.375,
+      "grad_norm_var": 3.2143229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3201,
+      "loss/crossentropy": 2.29172303378582,
+      "loss/hidden": 3.313671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19257347993552684,
+      "step": 22080
+    },
+    {
+      "epoch": 0.55225,
+      "grad_norm": 28.875,
+      "grad_norm_var": 4.649739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2483,
+      "loss/crossentropy": 1.9568318665027618,
+      "loss/hidden": 3.3109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17994072400033473,
+      "step": 22090
+    },
+    {
+      "epoch": 0.5525,
+      "grad_norm": 34.0,
+      "grad_norm_var": 4.592708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2302,
+      "loss/crossentropy": 2.148341643810272,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18362896144390106,
+      "step": 22100
+    },
+    {
+      "epoch": 0.55275,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.36640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2849,
+      "loss/crossentropy": 1.9461228162050248,
+      "loss/hidden": 3.508203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1895364910364151,
+      "step": 22110
+    },
+    {
+      "epoch": 0.553,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.9497395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4344,
+      "loss/crossentropy": 2.1932419329881667,
+      "loss/hidden": 3.39140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19293667376041412,
+      "step": 22120
+    },
+    {
+      "epoch": 0.55325,
+      "grad_norm": 29.75,
+      "grad_norm_var": 7.080143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3256,
+      "loss/crossentropy": 2.164583859592676,
+      "loss/hidden": 3.3296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1796217132359743,
+      "step": 22130
+    },
+    {
+      "epoch": 0.5535,
+      "grad_norm": 33.75,
+      "grad_norm_var": 4.024739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3128,
+      "loss/crossentropy": 2.0828025907278063,
+      "loss/hidden": 3.292578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17759974133223294,
+      "step": 22140
+    },
+    {
+      "epoch": 0.55375,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.9150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2276,
+      "loss/crossentropy": 1.9275901369750499,
+      "loss/hidden": 3.31171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16787631679326295,
+      "step": 22150
+    },
+    {
+      "epoch": 0.554,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.190559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.327,
+      "loss/crossentropy": 2.0888339787721635,
+      "loss/hidden": 3.300390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17749339248985052,
+      "step": 22160
+    },
+    {
+      "epoch": 0.55425,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.7129557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3324,
+      "loss/crossentropy": 1.9096436515450477,
+      "loss/hidden": 3.352734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16837584171444178,
+      "step": 22170
+    },
+    {
+      "epoch": 0.5545,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.463997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2819,
+      "loss/crossentropy": 2.0638594791293143,
+      "loss/hidden": 3.375390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1814063997939229,
+      "step": 22180
+    },
+    {
+      "epoch": 0.55475,
+      "grad_norm": 39.5,
+      "grad_norm_var": 7.167643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2706,
+      "loss/crossentropy": 2.1528440028429032,
+      "loss/hidden": 3.32421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19305985253304242,
+      "step": 22190
+    },
+    {
+      "epoch": 0.555,
+      "grad_norm": 34.5,
+      "grad_norm_var": 7.8875,
+      "learning_rate": 0.0001,
+      "loss": 7.3135,
+      "loss/crossentropy": 2.163337790966034,
+      "loss/hidden": 3.2890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17552390675991775,
+      "step": 22200
+    },
+    {
+      "epoch": 0.55525,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.520247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1905,
+      "loss/crossentropy": 2.0848088540136813,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1875115931034088,
+      "step": 22210
+    },
+    {
+      "epoch": 0.5555,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.845833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2472,
+      "loss/crossentropy": 2.200931230187416,
+      "loss/hidden": 3.366015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20207254495471716,
+      "step": 22220
+    },
+    {
+      "epoch": 0.55575,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.02265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3226,
+      "loss/crossentropy": 1.9815104238688945,
+      "loss/hidden": 3.331640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1747302796691656,
+      "step": 22230
+    },
+    {
+      "epoch": 0.556,
+      "grad_norm": 30.125,
+      "grad_norm_var": 4.238997395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3566,
+      "loss/crossentropy": 2.1646396666765213,
+      "loss/hidden": 3.48046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2087303439155221,
+      "step": 22240
+    },
+    {
+      "epoch": 0.55625,
+      "grad_norm": 31.875,
+      "grad_norm_var": 5.177018229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3546,
+      "loss/crossentropy": 2.0839816108345985,
+      "loss/hidden": 3.319921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18073585536330938,
+      "step": 22250
+    },
+    {
+      "epoch": 0.5565,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.723372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3226,
+      "loss/crossentropy": 2.0152421653270722,
+      "loss/hidden": 3.298828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17794121894985437,
+      "step": 22260
+    },
+    {
+      "epoch": 0.55675,
+      "grad_norm": 28.5,
+      "grad_norm_var": 8.039518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3145,
+      "loss/crossentropy": 2.0704862751066684,
+      "loss/hidden": 3.303125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17498300932347774,
+      "step": 22270
+    },
+    {
+      "epoch": 0.557,
+      "grad_norm": 31.125,
+      "grad_norm_var": 10.777083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2271,
+      "loss/crossentropy": 1.9257421031594277,
+      "loss/hidden": 3.4390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19484029989689589,
+      "step": 22280
+    },
+    {
+      "epoch": 0.55725,
+      "grad_norm": 30.5,
+      "grad_norm_var": 6.412239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.249,
+      "loss/crossentropy": 2.0617389529943466,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18348366860300303,
+      "step": 22290
+    },
+    {
+      "epoch": 0.5575,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.5582682291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3245,
+      "loss/crossentropy": 2.0664080172777175,
+      "loss/hidden": 3.355859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17801499664783477,
+      "step": 22300
+    },
+    {
+      "epoch": 0.55775,
+      "grad_norm": 28.375,
+      "grad_norm_var": 2.41015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2956,
+      "loss/crossentropy": 2.009297924488783,
+      "loss/hidden": 3.383984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18730297349393368,
+      "step": 22310
+    },
+    {
+      "epoch": 0.558,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.723893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3288,
+      "loss/crossentropy": 2.138043949007988,
+      "loss/hidden": 3.44375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19984424524009228,
+      "step": 22320
+    },
+    {
+      "epoch": 0.55825,
+      "grad_norm": 31.125,
+      "grad_norm_var": 9.493489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2771,
+      "loss/crossentropy": 2.038010062277317,
+      "loss/hidden": 3.310546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1933336803689599,
+      "step": 22330
+    },
+    {
+      "epoch": 0.5585,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.6884765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2633,
+      "loss/crossentropy": 2.2755305796861647,
+      "loss/hidden": 3.38046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19489618260413408,
+      "step": 22340
+    },
+    {
+      "epoch": 0.55875,
+      "grad_norm": 30.875,
+      "grad_norm_var": 51.78333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3329,
+      "loss/crossentropy": 2.179186634719372,
+      "loss/hidden": 3.3546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1874596353620291,
+      "step": 22350
+    },
+    {
+      "epoch": 0.559,
+      "grad_norm": 36.25,
+      "grad_norm_var": 38.72057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3535,
+      "loss/crossentropy": 1.9558674409985541,
+      "loss/hidden": 3.55546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19192036390304565,
+      "step": 22360
+    },
+    {
+      "epoch": 0.55925,
+      "grad_norm": 32.5,
+      "grad_norm_var": 52.962955729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4462,
+      "loss/crossentropy": 2.16518527418375,
+      "loss/hidden": 3.419140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1972449893131852,
+      "step": 22370
+    },
+    {
+      "epoch": 0.5595,
+      "grad_norm": 33.75,
+      "grad_norm_var": 41.524739583333336,
+      "learning_rate": 0.0001,
+      "loss": 7.2722,
+      "loss/crossentropy": 2.1358800739049912,
+      "loss/hidden": 3.298046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18384176325052975,
+      "step": 22380
+    },
+    {
+      "epoch": 0.55975,
+      "grad_norm": 41.5,
+      "grad_norm_var": 23.659830729166668,
+      "learning_rate": 0.0001,
+      "loss": 7.3707,
+      "loss/crossentropy": 2.090507461130619,
+      "loss/hidden": 3.310546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17735959570854903,
+      "step": 22390
+    },
+    {
+      "epoch": 0.56,
+      "grad_norm": 32.5,
+      "grad_norm_var": 15.437239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1671,
+      "loss/crossentropy": 2.0199001021683216,
+      "loss/hidden": 3.30546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.176216440834105,
+      "step": 22400
+    },
+    {
+      "epoch": 0.56025,
+      "grad_norm": 32.5,
+      "grad_norm_var": 12.668489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3144,
+      "loss/crossentropy": 2.0520263105630874,
+      "loss/hidden": 3.526953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.207816849835217,
+      "step": 22410
+    },
+    {
+      "epoch": 0.5605,
+      "grad_norm": 30.0,
+      "grad_norm_var": 10.243489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.251,
+      "loss/crossentropy": 2.191933420300484,
+      "loss/hidden": 3.293359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18148876912891865,
+      "step": 22420
+    },
+    {
+      "epoch": 0.56075,
+      "grad_norm": 27.625,
+      "grad_norm_var": 8.584830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2872,
+      "loss/crossentropy": 2.0942600347101687,
+      "loss/hidden": 3.2859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1781538650393486,
+      "step": 22430
+    },
+    {
+      "epoch": 0.561,
+      "grad_norm": 28.375,
+      "grad_norm_var": 30.66640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3342,
+      "loss/crossentropy": 2.0626624435186387,
+      "loss/hidden": 3.446484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20761814154684544,
+      "step": 22440
+    },
+    {
+      "epoch": 0.56125,
+      "grad_norm": 28.875,
+      "grad_norm_var": 5.724739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3023,
+      "loss/crossentropy": 2.153649944067001,
+      "loss/hidden": 3.321875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2057520892471075,
+      "step": 22450
+    },
+    {
+      "epoch": 0.5615,
+      "grad_norm": 30.0,
+      "grad_norm_var": 8.825455729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4168,
+      "loss/crossentropy": 2.1032363295555117,
+      "loss/hidden": 3.332421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17875799927860497,
+      "step": 22460
+    },
+    {
+      "epoch": 0.56175,
+      "grad_norm": 31.75,
+      "grad_norm_var": 6.794205729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2705,
+      "loss/crossentropy": 2.2260005325078964,
+      "loss/hidden": 3.319921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19685445744544267,
+      "step": 22470
+    },
+    {
+      "epoch": 0.562,
+      "grad_norm": 31.5,
+      "grad_norm_var": 4.7478515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2752,
+      "loss/crossentropy": 2.0784255638718605,
+      "loss/hidden": 3.301171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19737142305821181,
+      "step": 22480
+    },
+    {
+      "epoch": 0.56225,
+      "grad_norm": 34.75,
+      "grad_norm_var": 6.758072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3435,
+      "loss/crossentropy": 2.143521362543106,
+      "loss/hidden": 3.271875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17078035548329354,
+      "step": 22490
+    },
+    {
+      "epoch": 0.5625,
+      "grad_norm": 29.0,
+      "grad_norm_var": 7.628580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3417,
+      "loss/crossentropy": 2.0641577050089834,
+      "loss/hidden": 3.555078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1941977996379137,
+      "step": 22500
+    },
+    {
+      "epoch": 0.56275,
+      "grad_norm": 29.875,
+      "grad_norm_var": 7.066666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4512,
+      "loss/crossentropy": 2.1928183168172835,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19421984851360322,
+      "step": 22510
+    },
+    {
+      "epoch": 0.563,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.2030598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3443,
+      "loss/crossentropy": 2.068145313858986,
+      "loss/hidden": 3.468359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19339582826942206,
+      "step": 22520
+    },
+    {
+      "epoch": 0.56325,
+      "grad_norm": 31.25,
+      "grad_norm_var": 9.23515625,
+      "learning_rate": 0.0001,
+      "loss": 7.378,
+      "loss/crossentropy": 2.130034548044205,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19611325915902852,
+      "step": 22530
+    },
+    {
+      "epoch": 0.5635,
+      "grad_norm": 29.625,
+      "grad_norm_var": 11.611458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2822,
+      "loss/crossentropy": 1.9730725206434727,
+      "loss/hidden": 3.461328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2219389332458377,
+      "step": 22540
+    },
+    {
+      "epoch": 0.56375,
+      "grad_norm": 28.5,
+      "grad_norm_var": 3.5660807291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2197,
+      "loss/crossentropy": 2.150769717991352,
+      "loss/hidden": 3.38828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20134973097592593,
+      "step": 22550
+    },
+    {
+      "epoch": 0.564,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.5135416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2492,
+      "loss/crossentropy": 2.00289306640625,
+      "loss/hidden": 3.277734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17620294205844403,
+      "step": 22560
+    },
+    {
+      "epoch": 0.56425,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.3301432291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2268,
+      "loss/crossentropy": 2.106088588386774,
+      "loss/hidden": 3.379296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20606033485382796,
+      "step": 22570
+    },
+    {
+      "epoch": 0.5645,
+      "grad_norm": 28.125,
+      "grad_norm_var": 3.089583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2221,
+      "loss/crossentropy": 1.9868631184101104,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17915018759667872,
+      "step": 22580
+    },
+    {
+      "epoch": 0.56475,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.951041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4444,
+      "loss/crossentropy": 2.104344055056572,
+      "loss/hidden": 3.43359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19927312359213828,
+      "step": 22590
+    },
+    {
+      "epoch": 0.565,
+      "grad_norm": 28.875,
+      "grad_norm_var": 18.667643229166668,
+      "learning_rate": 0.0001,
+      "loss": 7.3948,
+      "loss/crossentropy": 2.129189969599247,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2167124640196562,
+      "step": 22600
+    },
+    {
+      "epoch": 0.56525,
+      "grad_norm": 31.625,
+      "grad_norm_var": 18.68515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3995,
+      "loss/crossentropy": 2.0350073277950287,
+      "loss/hidden": 3.36015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1885037964209914,
+      "step": 22610
+    },
+    {
+      "epoch": 0.5655,
+      "grad_norm": 35.25,
+      "grad_norm_var": 5.2375,
+      "learning_rate": 0.0001,
+      "loss": 7.1725,
+      "loss/crossentropy": 1.988342931866646,
+      "loss/hidden": 3.403515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1710897535085678,
+      "step": 22620
+    },
+    {
+      "epoch": 0.56575,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.592643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3575,
+      "loss/crossentropy": 2.1170627892017366,
+      "loss/hidden": 3.47109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20078443121165038,
+      "step": 22630
+    },
+    {
+      "epoch": 0.566,
+      "grad_norm": 29.25,
+      "grad_norm_var": 2.0160807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2549,
+      "loss/crossentropy": 1.9923384763300418,
+      "loss/hidden": 3.40859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2009210099466145,
+      "step": 22640
+    },
+    {
+      "epoch": 0.56625,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.9759765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3738,
+      "loss/crossentropy": 2.166072864830494,
+      "loss/hidden": 3.28828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1796887915581465,
+      "step": 22650
+    },
+    {
+      "epoch": 0.5665,
+      "grad_norm": 6274678784.0,
+      "grad_norm_var": 2.46072459061866e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3881,
+      "loss/crossentropy": 2.002020299434662,
+      "loss/hidden": 3.456640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2040075208991766,
+      "step": 22660
+    },
+    {
+      "epoch": 0.56675,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.460724590599052e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3138,
+      "loss/crossentropy": 2.0951256424188616,
+      "loss/hidden": 3.3375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17585320938378574,
+      "step": 22670
+    },
+    {
+      "epoch": 0.567,
+      "grad_norm": 31.25,
+      "grad_norm_var": 0.7947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.302,
+      "loss/crossentropy": 1.918744233250618,
+      "loss/hidden": 3.339453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1783786091953516,
+      "step": 22680
+    },
+    {
+      "epoch": 0.56725,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.9080729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3487,
+      "loss/crossentropy": 2.0890044644474983,
+      "loss/hidden": 3.31015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.186820235196501,
+      "step": 22690
+    },
+    {
+      "epoch": 0.5675,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.16015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3366,
+      "loss/crossentropy": 2.0049152970314026,
+      "loss/hidden": 3.364453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18571792878210544,
+      "step": 22700
+    },
+    {
+      "epoch": 0.56775,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.2119140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3629,
+      "loss/crossentropy": 2.188979035615921,
+      "loss/hidden": 3.303515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18791910037398338,
+      "step": 22710
+    },
+    {
+      "epoch": 0.568,
+      "grad_norm": 28.125,
+      "grad_norm_var": 6.873958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2721,
+      "loss/crossentropy": 1.9086629442870617,
+      "loss/hidden": 3.314453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1676910617388785,
+      "step": 22720
+    },
+    {
+      "epoch": 0.56825,
+      "grad_norm": 29.375,
+      "grad_norm_var": 4.77890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2104,
+      "loss/crossentropy": 1.9729705944657325,
+      "loss/hidden": 3.31484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18355059083551167,
+      "step": 22730
+    },
+    {
+      "epoch": 0.5685,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.0416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3276,
+      "loss/crossentropy": 1.983503246307373,
+      "loss/hidden": 3.3828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1911757795140147,
+      "step": 22740
+    },
+    {
+      "epoch": 0.56875,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.209830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3981,
+      "loss/crossentropy": 2.0647961035370828,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19169409601017834,
+      "step": 22750
+    },
+    {
+      "epoch": 0.569,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.794205729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3512,
+      "loss/crossentropy": 2.0892448581755163,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18739675851538778,
+      "step": 22760
+    },
+    {
+      "epoch": 0.56925,
+      "grad_norm": 30.5,
+      "grad_norm_var": 4.69375,
+      "learning_rate": 0.0001,
+      "loss": 7.3437,
+      "loss/crossentropy": 2.052200788259506,
+      "loss/hidden": 3.51875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19390411972999572,
+      "step": 22770
+    },
+    {
+      "epoch": 0.5695,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.9322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2241,
+      "loss/crossentropy": 2.2473602324724196,
+      "loss/hidden": 3.26875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18824474215507508,
+      "step": 22780
+    },
+    {
+      "epoch": 0.56975,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.1973307291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4546,
+      "loss/crossentropy": 2.2174068212509157,
+      "loss/hidden": 3.44765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2089033018797636,
+      "step": 22790
+    },
+    {
+      "epoch": 0.57,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.8775390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2814,
+      "loss/crossentropy": 2.098750987648964,
+      "loss/hidden": 3.439453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18969560451805592,
+      "step": 22800
+    },
+    {
+      "epoch": 0.57025,
+      "grad_norm": 29.0,
+      "grad_norm_var": 4.448958333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3282,
+      "loss/crossentropy": 2.2710734084248543,
+      "loss/hidden": 3.3,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18953933827579023,
+      "step": 22810
+    },
+    {
+      "epoch": 0.5705,
+      "grad_norm": 27.75,
+      "grad_norm_var": 10.401041666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2824,
+      "loss/crossentropy": 2.08201609402895,
+      "loss/hidden": 3.215234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16761352298781276,
+      "step": 22820
+    },
+    {
+      "epoch": 0.57075,
+      "grad_norm": 28.75,
+      "grad_norm_var": 12.773893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4861,
+      "loss/crossentropy": 2.1187492191791533,
+      "loss/hidden": 3.484765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19788370802998542,
+      "step": 22830
+    },
+    {
+      "epoch": 0.571,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.25625,
+      "learning_rate": 0.0001,
+      "loss": 7.34,
+      "loss/crossentropy": 2.141358491778374,
+      "loss/hidden": 3.31328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17862968128174544,
+      "step": 22840
+    },
+    {
+      "epoch": 0.57125,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.4372395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3556,
+      "loss/crossentropy": 2.159883065521717,
+      "loss/hidden": 3.427734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20434525702148676,
+      "step": 22850
+    },
+    {
+      "epoch": 0.5715,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.6311848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3261,
+      "loss/crossentropy": 2.0141698867082596,
+      "loss/hidden": 3.323046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17547854110598565,
+      "step": 22860
+    },
+    {
+      "epoch": 0.57175,
+      "grad_norm": 27.875,
+      "grad_norm_var": 1.8858723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1736,
+      "loss/crossentropy": 2.070405501127243,
+      "loss/hidden": 3.28515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17701356951147318,
+      "step": 22870
+    },
+    {
+      "epoch": 0.572,
+      "grad_norm": 33.75,
+      "grad_norm_var": 57.8150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3726,
+      "loss/crossentropy": 2.1324386440217493,
+      "loss/hidden": 3.46328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20903213806450366,
+      "step": 22880
+    },
+    {
+      "epoch": 0.57225,
+      "grad_norm": 30.125,
+      "grad_norm_var": 57.978580729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2264,
+      "loss/crossentropy": 2.056166473031044,
+      "loss/hidden": 3.353515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19125475976616144,
+      "step": 22890
+    },
+    {
+      "epoch": 0.5725,
+      "grad_norm": 29.0,
+      "grad_norm_var": 2.1166015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3529,
+      "loss/crossentropy": 2.0231011360883713,
+      "loss/hidden": 3.454296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19594160905107855,
+      "step": 22900
+    },
+    {
+      "epoch": 0.57275,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.2059895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2611,
+      "loss/crossentropy": 2.1009394943714144,
+      "loss/hidden": 3.421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20553287640213966,
+      "step": 22910
+    },
+    {
+      "epoch": 0.573,
+      "grad_norm": 27.625,
+      "grad_norm_var": 2.99765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2733,
+      "loss/crossentropy": 2.1342287354171274,
+      "loss/hidden": 3.35546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19136658031493425,
+      "step": 22920
+    },
+    {
+      "epoch": 0.57325,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.39765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2991,
+      "loss/crossentropy": 2.0322689548134805,
+      "loss/hidden": 3.269921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17143328655511142,
+      "step": 22930
+    },
+    {
+      "epoch": 0.5735,
+      "grad_norm": 29.0,
+      "grad_norm_var": 1.94765625,
+      "learning_rate": 0.0001,
+      "loss": 7.278,
+      "loss/crossentropy": 2.1039293974637987,
+      "loss/hidden": 3.244140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1707957200706005,
+      "step": 22940
+    },
+    {
+      "epoch": 0.57375,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.2108723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2094,
+      "loss/crossentropy": 2.0728151589632033,
+      "loss/hidden": 3.294921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17216427512466909,
+      "step": 22950
+    },
+    {
+      "epoch": 0.574,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.4358723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3065,
+      "loss/crossentropy": 2.161167304217815,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18787571005523204,
+      "step": 22960
+    },
+    {
+      "epoch": 0.57425,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.8697916666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3397,
+      "loss/crossentropy": 2.104795481264591,
+      "loss/hidden": 3.309765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1772844972088933,
+      "step": 22970
+    },
+    {
+      "epoch": 0.5745,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.2395833333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3098,
+      "loss/crossentropy": 2.170831048488617,
+      "loss/hidden": 3.323046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1898819526657462,
+      "step": 22980
+    },
+    {
+      "epoch": 0.57475,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.8598307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2474,
+      "loss/crossentropy": 2.072402949631214,
+      "loss/hidden": 3.4015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20032873265445234,
+      "step": 22990
+    },
+    {
+      "epoch": 0.575,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.0864583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1527,
+      "loss/crossentropy": 1.9359148509800435,
+      "loss/hidden": 3.36953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17023193156346678,
+      "step": 23000
+    },
+    {
+      "epoch": 0.57525,
+      "grad_norm": 31.0,
+      "grad_norm_var": 10.963541666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3357,
+      "loss/crossentropy": 2.1682570412755013,
+      "loss/hidden": 3.41015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20298131536692382,
+      "step": 23010
+    },
+    {
+      "epoch": 0.5755,
+      "grad_norm": 33.25,
+      "grad_norm_var": 9.428125,
+      "learning_rate": 0.0001,
+      "loss": 7.3109,
+      "loss/crossentropy": 2.045479938387871,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18103279825299978,
+      "step": 23020
+    },
+    {
+      "epoch": 0.57575,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.339583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3223,
+      "loss/crossentropy": 2.197012846916914,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1997917301952839,
+      "step": 23030
+    },
+    {
+      "epoch": 0.576,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.3955729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.328,
+      "loss/crossentropy": 2.0420756086707117,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19402425028383732,
+      "step": 23040
+    },
+    {
+      "epoch": 0.57625,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.7926432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3225,
+      "loss/crossentropy": 2.227956184744835,
+      "loss/hidden": 3.33359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20268909279257058,
+      "step": 23050
+    },
+    {
+      "epoch": 0.5765,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.319205729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3548,
+      "loss/crossentropy": 2.03707487732172,
+      "loss/hidden": 3.4078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18702135644853116,
+      "step": 23060
+    },
+    {
+      "epoch": 0.57675,
+      "grad_norm": 31.5,
+      "grad_norm_var": 3.103261589190476e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3001,
+      "loss/crossentropy": 2.2075371608138084,
+      "loss/hidden": 3.287109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18566503124311567,
+      "step": 23070
+    },
+    {
+      "epoch": 0.577,
+      "grad_norm": 36.25,
+      "grad_norm_var": 4.934309895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.244,
+      "loss/crossentropy": 2.0198142036795614,
+      "loss/hidden": 3.377734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18854798339307308,
+      "step": 23080
+    },
+    {
+      "epoch": 0.57725,
+      "grad_norm": 32.5,
+      "grad_norm_var": 6.170768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3501,
+      "loss/crossentropy": 2.1881318643689154,
+      "loss/hidden": 3.32890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19292537569999696,
+      "step": 23090
+    },
+    {
+      "epoch": 0.5775,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.0249348958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2021,
+      "loss/crossentropy": 2.0514975383877756,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1834509514272213,
+      "step": 23100
+    },
+    {
+      "epoch": 0.57775,
+      "grad_norm": 27.875,
+      "grad_norm_var": 2.409375,
+      "learning_rate": 0.0001,
+      "loss": 7.2787,
+      "loss/crossentropy": 1.8552511997520924,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.172595002502203,
+      "step": 23110
+    },
+    {
+      "epoch": 0.578,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.903580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3178,
+      "loss/crossentropy": 2.0356574684381483,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1825650293380022,
+      "step": 23120
+    },
+    {
+      "epoch": 0.57825,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.1791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2816,
+      "loss/crossentropy": 1.8349962562322617,
+      "loss/hidden": 3.39375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18128557410091162,
+      "step": 23130
+    },
+    {
+      "epoch": 0.5785,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.8479166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.435,
+      "loss/crossentropy": 1.9793405070900918,
+      "loss/hidden": 3.430078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18810991439968347,
+      "step": 23140
+    },
+    {
+      "epoch": 0.57875,
+      "grad_norm": 27.75,
+      "grad_norm_var": 1.3958333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3526,
+      "loss/crossentropy": 2.2002530977129937,
+      "loss/hidden": 3.287890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1851093016564846,
+      "step": 23150
+    },
+    {
+      "epoch": 0.579,
+      "grad_norm": 30.375,
+      "grad_norm_var": 9.785872395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2364,
+      "loss/crossentropy": 1.8952147908508779,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17190355882048608,
+      "step": 23160
+    },
+    {
+      "epoch": 0.57925,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.5212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3738,
+      "loss/crossentropy": 2.0228386133909226,
+      "loss/hidden": 3.455859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20162178240716458,
+      "step": 23170
+    },
+    {
+      "epoch": 0.5795,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.876822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3244,
+      "loss/crossentropy": 2.1140632264316084,
+      "loss/hidden": 3.41640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18904146291315554,
+      "step": 23180
+    },
+    {
+      "epoch": 0.57975,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.46015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3871,
+      "loss/crossentropy": 2.046115532517433,
+      "loss/hidden": 3.364453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1814727798104286,
+      "step": 23190
+    },
+    {
+      "epoch": 0.58,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.8806640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2065,
+      "loss/crossentropy": 2.09497309550643,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18063673991709947,
+      "step": 23200
+    },
+    {
+      "epoch": 0.58025,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.7184895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2516,
+      "loss/crossentropy": 2.120602674782276,
+      "loss/hidden": 3.3296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17760662753134965,
+      "step": 23210
+    },
+    {
+      "epoch": 0.5805,
+      "grad_norm": 29.375,
+      "grad_norm_var": 12.099739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3189,
+      "loss/crossentropy": 1.8746614530682564,
+      "loss/hidden": 3.321875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17366855386644603,
+      "step": 23220
+    },
+    {
+      "epoch": 0.58075,
+      "grad_norm": 34.0,
+      "grad_norm_var": 8.370572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.219,
+      "loss/crossentropy": 1.9041770923882724,
+      "loss/hidden": 3.337890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17128421002998948,
+      "step": 23230
+    },
+    {
+      "epoch": 0.581,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.69765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2582,
+      "loss/crossentropy": 2.141513818502426,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19048179611563681,
+      "step": 23240
+    },
+    {
+      "epoch": 0.58125,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.5434895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2658,
+      "loss/crossentropy": 2.0868393763899804,
+      "loss/hidden": 3.388671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19045372307300568,
+      "step": 23250
+    },
+    {
+      "epoch": 0.5815,
+      "grad_norm": 29.375,
+      "grad_norm_var": 5.077018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2014,
+      "loss/crossentropy": 2.101007029414177,
+      "loss/hidden": 3.36328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1854638984426856,
+      "step": 23260
+    },
+    {
+      "epoch": 0.58175,
+      "grad_norm": 31.125,
+      "grad_norm_var": 218.71920572916667,
+      "learning_rate": 0.0001,
+      "loss": 7.3693,
+      "loss/crossentropy": 1.9861895442008972,
+      "loss/hidden": 3.498046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20660532703623175,
+      "step": 23270
+    },
+    {
+      "epoch": 0.582,
+      "grad_norm": 28.625,
+      "grad_norm_var": 225.99140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2359,
+      "loss/crossentropy": 1.9705702036619186,
+      "loss/hidden": 3.25625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1714084181934595,
+      "step": 23280
+    },
+    {
+      "epoch": 0.58225,
+      "grad_norm": 33.25,
+      "grad_norm_var": 1.9921223958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2894,
+      "loss/crossentropy": 1.9414188079535961,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18114634826779366,
+      "step": 23290
+    },
+    {
+      "epoch": 0.5825,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.2875,
+      "learning_rate": 0.0001,
+      "loss": 7.1979,
+      "loss/crossentropy": 2.05298957452178,
+      "loss/hidden": 3.316015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18091321904212238,
+      "step": 23300
+    },
+    {
+      "epoch": 0.58275,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.949934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2295,
+      "loss/crossentropy": 2.0978086851537228,
+      "loss/hidden": 3.2453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17559153838083147,
+      "step": 23310
+    },
+    {
+      "epoch": 0.583,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.1143229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3858,
+      "loss/crossentropy": 2.1657110869884493,
+      "loss/hidden": 3.28515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17203837148845197,
+      "step": 23320
+    },
+    {
+      "epoch": 0.58325,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.0768229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3152,
+      "loss/crossentropy": 2.141623441874981,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19364709984511136,
+      "step": 23330
+    },
+    {
+      "epoch": 0.5835,
+      "grad_norm": 30.625,
+      "grad_norm_var": 4.070572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2853,
+      "loss/crossentropy": 2.1005776453763247,
+      "loss/hidden": 3.31640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1824392621871084,
+      "step": 23340
+    },
+    {
+      "epoch": 0.58375,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.6770833333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2691,
+      "loss/crossentropy": 2.182777139544487,
+      "loss/hidden": 3.339453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19071639832109213,
+      "step": 23350
+    },
+    {
+      "epoch": 0.584,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.8061848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2273,
+      "loss/crossentropy": 2.005091509968042,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17102911770343782,
+      "step": 23360
+    },
+    {
+      "epoch": 0.58425,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.6129557291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3946,
+      "loss/crossentropy": 1.958120694756508,
+      "loss/hidden": 3.47890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20584547519683838,
+      "step": 23370
+    },
+    {
+      "epoch": 0.5845,
+      "grad_norm": 33.75,
+      "grad_norm_var": 3.59765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4152,
+      "loss/crossentropy": 2.119162234663963,
+      "loss/hidden": 3.400390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19411879703402518,
+      "step": 23380
+    },
+    {
+      "epoch": 0.58475,
+      "grad_norm": 30.75,
+      "grad_norm_var": 4.620572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3097,
+      "loss/crossentropy": 2.1464238077402116,
+      "loss/hidden": 3.4796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19650665801018477,
+      "step": 23390
+    },
+    {
+      "epoch": 0.585,
+      "grad_norm": 30.0,
+      "grad_norm_var": 20.2478515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4,
+      "loss/crossentropy": 2.039510624855757,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18683098405599594,
+      "step": 23400
+    },
+    {
+      "epoch": 0.58525,
+      "grad_norm": 32.25,
+      "grad_norm_var": 41.15,
+      "learning_rate": 0.0001,
+      "loss": 7.3072,
+      "loss/crossentropy": 2.182946425676346,
+      "loss/hidden": 3.33359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19105008877813817,
+      "step": 23410
+    },
+    {
+      "epoch": 0.5855,
+      "grad_norm": 35.25,
+      "grad_norm_var": 4.7994140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3148,
+      "loss/crossentropy": 2.0956947550177576,
+      "loss/hidden": 3.43828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20151016414165496,
+      "step": 23420
+    },
+    {
+      "epoch": 0.58575,
+      "grad_norm": 34.5,
+      "grad_norm_var": 4.302018229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3171,
+      "loss/crossentropy": 2.0624319493770598,
+      "loss/hidden": 3.38359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17940568849444388,
+      "step": 23430
+    },
+    {
+      "epoch": 0.586,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.205208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2999,
+      "loss/crossentropy": 2.1612589955329895,
+      "loss/hidden": 3.279296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18243170958012342,
+      "step": 23440
+    },
+    {
+      "epoch": 0.58625,
+      "grad_norm": 43.0,
+      "grad_norm_var": 16.95,
+      "learning_rate": 0.0001,
+      "loss": 7.3768,
+      "loss/crossentropy": 2.119496448338032,
+      "loss/hidden": 3.4359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21236374396830798,
+      "step": 23450
+    },
+    {
+      "epoch": 0.5865,
+      "grad_norm": 30.5,
+      "grad_norm_var": 16.997330729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2713,
+      "loss/crossentropy": 2.226336379349232,
+      "loss/hidden": 3.37109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18839403800666332,
+      "step": 23460
+    },
+    {
+      "epoch": 0.58675,
+      "grad_norm": 30.25,
+      "grad_norm_var": 5.237239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1908,
+      "loss/crossentropy": 2.0912389412522314,
+      "loss/hidden": 3.446875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18238936755806207,
+      "step": 23470
+    },
+    {
+      "epoch": 0.587,
+      "grad_norm": 36.75,
+      "grad_norm_var": 5.435416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1305,
+      "loss/crossentropy": 2.014119653403759,
+      "loss/hidden": 3.308203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16768602933734655,
+      "step": 23480
+    },
+    {
+      "epoch": 0.58725,
+      "grad_norm": 31.875,
+      "grad_norm_var": 11.8509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2135,
+      "loss/crossentropy": 2.193871647119522,
+      "loss/hidden": 3.266796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1774727776646614,
+      "step": 23490
+    },
+    {
+      "epoch": 0.5875,
+      "grad_norm": 30.5,
+      "grad_norm_var": 8.875,
+      "learning_rate": 0.0001,
+      "loss": 7.2883,
+      "loss/crossentropy": 1.9861755840480328,
+      "loss/hidden": 3.36953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18203947991132735,
+      "step": 23500
+    },
+    {
+      "epoch": 0.58775,
+      "grad_norm": 34.25,
+      "grad_norm_var": 3.1697265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3764,
+      "loss/crossentropy": 2.1041073873639107,
+      "loss/hidden": 3.298046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1823105582036078,
+      "step": 23510
+    },
+    {
+      "epoch": 0.588,
+      "grad_norm": 28.75,
+      "grad_norm_var": 6.472916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2363,
+      "loss/crossentropy": 2.0386205092072487,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18798258919268845,
+      "step": 23520
+    },
+    {
+      "epoch": 0.58825,
+      "grad_norm": 29.5,
+      "grad_norm_var": 4.173958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2541,
+      "loss/crossentropy": 1.9665139004588128,
+      "loss/hidden": 3.41640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19156214147806166,
+      "step": 23530
+    },
+    {
+      "epoch": 0.5885,
+      "grad_norm": 28.375,
+      "grad_norm_var": 2.6264973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1821,
+      "loss/crossentropy": 2.100816609710455,
+      "loss/hidden": 3.336328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18041261453181506,
+      "step": 23540
+    },
+    {
+      "epoch": 0.58875,
+      "grad_norm": 28.25,
+      "grad_norm_var": 6.0681640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2194,
+      "loss/crossentropy": 2.160541406273842,
+      "loss/hidden": 3.279296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18264771308749914,
+      "step": 23550
+    },
+    {
+      "epoch": 0.589,
+      "grad_norm": 41.5,
+      "grad_norm_var": 13.784309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3172,
+      "loss/crossentropy": 2.0907914683222772,
+      "loss/hidden": 3.416796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19286628402769565,
+      "step": 23560
+    },
+    {
+      "epoch": 0.58925,
+      "grad_norm": 26.25,
+      "grad_norm_var": 33.94576822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.0992,
+      "loss/crossentropy": 2.035450255870819,
+      "loss/hidden": 3.325,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17769902311265467,
+      "step": 23570
+    },
+    {
+      "epoch": 0.5895,
+      "grad_norm": 29.375,
+      "grad_norm_var": 31.033268229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2749,
+      "loss/crossentropy": 2.011720988154411,
+      "loss/hidden": 3.331640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18213979490101337,
+      "step": 23580
+    },
+    {
+      "epoch": 0.58975,
+      "grad_norm": 28.0,
+      "grad_norm_var": 39.16451822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.2706,
+      "loss/crossentropy": 2.0319022722542286,
+      "loss/hidden": 3.330859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1784214310348034,
+      "step": 23590
+    },
+    {
+      "epoch": 0.59,
+      "grad_norm": 43.25,
+      "grad_norm_var": 43.842122395833336,
+      "learning_rate": 0.0001,
+      "loss": 7.2368,
+      "loss/crossentropy": 2.042397302389145,
+      "loss/hidden": 3.358984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17687461860477924,
+      "step": 23600
+    },
+    {
+      "epoch": 0.59025,
+      "grad_norm": 32.75,
+      "grad_norm_var": 13.026822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2755,
+      "loss/crossentropy": 1.9421588897705078,
+      "loss/hidden": 3.371875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1788268475793302,
+      "step": 23610
+    },
+    {
+      "epoch": 0.5905,
+      "grad_norm": 32.25,
+      "grad_norm_var": 7.690625,
+      "learning_rate": 0.0001,
+      "loss": 7.1561,
+      "loss/crossentropy": 2.0542794331908225,
+      "loss/hidden": 3.26796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16889538057148457,
+      "step": 23620
+    },
+    {
+      "epoch": 0.59075,
+      "grad_norm": 33.5,
+      "grad_norm_var": 10.685416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1786,
+      "loss/crossentropy": 2.0298443526029586,
+      "loss/hidden": 3.36171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1826344771310687,
+      "step": 23630
+    },
+    {
+      "epoch": 0.591,
+      "grad_norm": 39.5,
+      "grad_norm_var": 11.2806640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2082,
+      "loss/crossentropy": 2.119662294536829,
+      "loss/hidden": 3.30234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18428312726318835,
+      "step": 23640
+    },
+    {
+      "epoch": 0.59125,
+      "grad_norm": 29.75,
+      "grad_norm_var": 8.54765625,
+      "learning_rate": 0.0001,
+      "loss": 7.1784,
+      "loss/crossentropy": 1.9878482483327389,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18119903113692998,
+      "step": 23650
+    },
+    {
+      "epoch": 0.5915,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.0336566844344238e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3489,
+      "loss/crossentropy": 1.8514883562922477,
+      "loss/hidden": 3.5328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21164753548800946,
+      "step": 23660
+    },
+    {
+      "epoch": 0.59175,
+      "grad_norm": 31.0,
+      "grad_norm_var": 11.344791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3203,
+      "loss/crossentropy": 2.245069167017937,
+      "loss/hidden": 3.361328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17927178423851728,
+      "step": 23670
+    },
+    {
+      "epoch": 0.592,
+      "grad_norm": 29.0,
+      "grad_norm_var": 413.53020833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2885,
+      "loss/crossentropy": 2.0498727947473525,
+      "loss/hidden": 3.408203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18411201229318977,
+      "step": 23680
+    },
+    {
+      "epoch": 0.59225,
+      "grad_norm": 27.375,
+      "grad_norm_var": 25.974934895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.1401,
+      "loss/crossentropy": 1.9942969411611557,
+      "loss/hidden": 3.302734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16860720347613095,
+      "step": 23690
+    },
+    {
+      "epoch": 0.5925,
+      "grad_norm": 29.875,
+      "grad_norm_var": 8.026822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2296,
+      "loss/crossentropy": 2.0294051103293897,
+      "loss/hidden": 3.26875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18335335925221444,
+      "step": 23700
+    },
+    {
+      "epoch": 0.59275,
+      "grad_norm": 28.5,
+      "grad_norm_var": 9.404166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.272,
+      "loss/crossentropy": 2.1226597487926484,
+      "loss/hidden": 3.325,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17993064355105162,
+      "step": 23710
+    },
+    {
+      "epoch": 0.593,
+      "grad_norm": 27.125,
+      "grad_norm_var": 14.909375,
+      "learning_rate": 0.0001,
+      "loss": 7.2318,
+      "loss/crossentropy": 2.154828441143036,
+      "loss/hidden": 3.29375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18905305583029985,
+      "step": 23720
+    },
+    {
+      "epoch": 0.59325,
+      "grad_norm": 31.0,
+      "grad_norm_var": 9.435416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2639,
+      "loss/crossentropy": 2.0010774195194245,
+      "loss/hidden": 3.313671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1794521464034915,
+      "step": 23730
+    },
+    {
+      "epoch": 0.5935,
+      "grad_norm": 31.875,
+      "grad_norm_var": 7.265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2917,
+      "loss/crossentropy": 2.1218893982470037,
+      "loss/hidden": 3.341796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17976281177252532,
+      "step": 23740
+    },
+    {
+      "epoch": 0.59375,
+      "grad_norm": 29.875,
+      "grad_norm_var": 5.7837890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2556,
+      "loss/crossentropy": 2.0276944383978845,
+      "loss/hidden": 3.320703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1864795895293355,
+      "step": 23750
+    },
+    {
+      "epoch": 0.594,
+      "grad_norm": 34.0,
+      "grad_norm_var": 7.016666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2928,
+      "loss/crossentropy": 1.929387104511261,
+      "loss/hidden": 3.30703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17239041812717915,
+      "step": 23760
+    },
+    {
+      "epoch": 0.59425,
+      "grad_norm": 31.0,
+      "grad_norm_var": 7.468489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3058,
+      "loss/crossentropy": 2.0659641668200495,
+      "loss/hidden": 3.361328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1967785032466054,
+      "step": 23770
+    },
+    {
+      "epoch": 0.5945,
+      "grad_norm": 28.375,
+      "grad_norm_var": 5.722330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2596,
+      "loss/crossentropy": 2.0173033215105534,
+      "loss/hidden": 3.3078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17501826863735914,
+      "step": 23780
+    },
+    {
+      "epoch": 0.59475,
+      "grad_norm": 27.75,
+      "grad_norm_var": 5.6806640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3387,
+      "loss/crossentropy": 2.063326106220484,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.181632025167346,
+      "step": 23790
+    },
+    {
+      "epoch": 0.595,
+      "grad_norm": 29.0,
+      "grad_norm_var": 5.265559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3309,
+      "loss/crossentropy": 2.200699530541897,
+      "loss/hidden": 3.30859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1817481342703104,
+      "step": 23800
+    },
+    {
+      "epoch": 0.59525,
+      "grad_norm": 40.5,
+      "grad_norm_var": 3.8861138646518113e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3693,
+      "loss/crossentropy": 2.1953633323311808,
+      "loss/hidden": 3.3140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18916522301733493,
+      "step": 23810
+    },
+    {
+      "epoch": 0.5955,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.1892317594268467e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.1476,
+      "loss/crossentropy": 2.100045497715473,
+      "loss/hidden": 3.364453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18449472673237324,
+      "step": 23820
+    },
+    {
+      "epoch": 0.59575,
+      "grad_norm": 29.875,
+      "grad_norm_var": 7.113541666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2694,
+      "loss/crossentropy": 2.13639819920063,
+      "loss/hidden": 3.260546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18081072345376015,
+      "step": 23830
+    },
+    {
+      "epoch": 0.596,
+      "grad_norm": 32.75,
+      "grad_norm_var": 10.190625,
+      "learning_rate": 0.0001,
+      "loss": 7.4218,
+      "loss/crossentropy": 2.0917629063129426,
+      "loss/hidden": 3.3,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20855254549533128,
+      "step": 23840
+    },
+    {
+      "epoch": 0.59625,
+      "grad_norm": 35.25,
+      "grad_norm_var": 7.001822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3759,
+      "loss/crossentropy": 2.1656418085098266,
+      "loss/hidden": 3.341015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1982706617563963,
+      "step": 23850
+    },
+    {
+      "epoch": 0.5965,
+      "grad_norm": 30.625,
+      "grad_norm_var": 14.1806640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3419,
+      "loss/crossentropy": 1.978961955010891,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19264845587313176,
+      "step": 23860
+    },
+    {
+      "epoch": 0.59675,
+      "grad_norm": 28.75,
+      "grad_norm_var": 9.461458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1425,
+      "loss/crossentropy": 2.064087450504303,
+      "loss/hidden": 3.2859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18434563372284174,
+      "step": 23870
+    },
+    {
+      "epoch": 0.597,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.3979166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2846,
+      "loss/crossentropy": 2.2514883071184157,
+      "loss/hidden": 3.32578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1978050796315074,
+      "step": 23880
+    },
+    {
+      "epoch": 0.59725,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.5010416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3878,
+      "loss/crossentropy": 2.1330555945634844,
+      "loss/hidden": 3.3546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1861652435734868,
+      "step": 23890
+    },
+    {
+      "epoch": 0.5975,
+      "grad_norm": 33.25,
+      "grad_norm_var": 5.234309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2809,
+      "loss/crossentropy": 2.15719972550869,
+      "loss/hidden": 3.323046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18762489072978497,
+      "step": 23900
+    },
+    {
+      "epoch": 0.59775,
+      "grad_norm": 28.125,
+      "grad_norm_var": 9.90625,
+      "learning_rate": 0.0001,
+      "loss": 7.3282,
+      "loss/crossentropy": 2.2501418024301527,
+      "loss/hidden": 3.288671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18364315424114466,
+      "step": 23910
+    },
+    {
+      "epoch": 0.598,
+      "grad_norm": 7885291520.0,
+      "grad_norm_var": 3.8861138654567685e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4436,
+      "loss/crossentropy": 1.849592723697424,
+      "loss/hidden": 3.33828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1732720138505101,
+      "step": 23920
+    },
+    {
+      "epoch": 0.59825,
+      "grad_norm": 35.25,
+      "grad_norm_var": 3.886113865095359e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3624,
+      "loss/crossentropy": 2.0290102303028106,
+      "loss/hidden": 3.36640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19378960439935328,
+      "step": 23930
+    },
+    {
+      "epoch": 0.5985,
+      "grad_norm": 28.25,
+      "grad_norm_var": 4.639322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2436,
+      "loss/crossentropy": 1.921845416724682,
+      "loss/hidden": 3.34140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19508706871420145,
+      "step": 23940
+    },
+    {
+      "epoch": 0.59875,
+      "grad_norm": 33.25,
+      "grad_norm_var": 3.9879557291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2391,
+      "loss/crossentropy": 2.113616520166397,
+      "loss/hidden": 3.190234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1806779010221362,
+      "step": 23950
+    },
+    {
+      "epoch": 0.599,
+      "grad_norm": 31.5,
+      "grad_norm_var": 4.765559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2259,
+      "loss/crossentropy": 2.0889094918966293,
+      "loss/hidden": 3.372265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19335521198809147,
+      "step": 23960
+    },
+    {
+      "epoch": 0.59925,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.7759765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3169,
+      "loss/crossentropy": 2.0757755756378176,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18353872746229172,
+      "step": 23970
+    },
+    {
+      "epoch": 0.5995,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.417122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.354,
+      "loss/crossentropy": 2.151854282617569,
+      "loss/hidden": 3.40234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18792435489594936,
+      "step": 23980
+    },
+    {
+      "epoch": 0.59975,
+      "grad_norm": 30.75,
+      "grad_norm_var": 15.67890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3461,
+      "loss/crossentropy": 1.9950793728232383,
+      "loss/hidden": 3.3421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20234323088079692,
+      "step": 23990
+    },
+    {
+      "epoch": 0.6,
+      "grad_norm": 29.875,
+      "grad_norm_var": 20.184309895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2263,
+      "loss/crossentropy": 1.9956952631473541,
+      "loss/hidden": 3.4046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18217481905594468,
+      "step": 24000
+    },
+    {
+      "epoch": 0.60025,
+      "grad_norm": 38.0,
+      "grad_norm_var": 9.999739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3267,
+      "loss/crossentropy": 2.0959785729646683,
+      "loss/hidden": 3.436328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1935113796964288,
+      "step": 24010
+    },
+    {
+      "epoch": 0.6005,
+      "grad_norm": 32.75,
+      "grad_norm_var": 5.863997395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2052,
+      "loss/crossentropy": 2.023172339051962,
+      "loss/hidden": 3.32890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17665521912276744,
+      "step": 24020
+    },
+    {
+      "epoch": 0.60075,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.0645182291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3997,
+      "loss/crossentropy": 2.044221018254757,
+      "loss/hidden": 3.3171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17454885710030793,
+      "step": 24030
+    },
+    {
+      "epoch": 0.601,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.1791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4056,
+      "loss/crossentropy": 2.017246203124523,
+      "loss/hidden": 3.40078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19033315852284433,
+      "step": 24040
+    },
+    {
+      "epoch": 0.60125,
+      "grad_norm": 34.0,
+      "grad_norm_var": 7304.769791666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3027,
+      "loss/crossentropy": 2.0873162485659122,
+      "loss/hidden": 3.305078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1803006211295724,
+      "step": 24050
+    },
+    {
+      "epoch": 0.6015,
+      "grad_norm": 42.0,
+      "grad_norm_var": 7222.07265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3271,
+      "loss/crossentropy": 2.089706966280937,
+      "loss/hidden": 3.2859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17721448224037886,
+      "step": 24060
+    },
+    {
+      "epoch": 0.60175,
+      "grad_norm": 28.5,
+      "grad_norm_var": 17.274934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2893,
+      "loss/crossentropy": 2.1399046301841738,
+      "loss/hidden": 3.32421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17965221870690584,
+      "step": 24070
+    },
+    {
+      "epoch": 0.602,
+      "grad_norm": 27.875,
+      "grad_norm_var": 3.4098307291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2298,
+      "loss/crossentropy": 2.232226923108101,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17638658490031958,
+      "step": 24080
+    },
+    {
+      "epoch": 0.60225,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.467643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2752,
+      "loss/crossentropy": 2.0909645952284337,
+      "loss/hidden": 3.324609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18019966538995505,
+      "step": 24090
+    },
+    {
+      "epoch": 0.6025,
+      "grad_norm": 33.25,
+      "grad_norm_var": 4.315559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1836,
+      "loss/crossentropy": 2.043725144863129,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1770002892240882,
+      "step": 24100
+    },
+    {
+      "epoch": 0.60275,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.894791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2902,
+      "loss/crossentropy": 2.09500589966774,
+      "loss/hidden": 3.276953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18369378615170717,
+      "step": 24110
+    },
+    {
+      "epoch": 0.603,
+      "grad_norm": 27.125,
+      "grad_norm_var": 5.116080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3256,
+      "loss/crossentropy": 1.9551431708037854,
+      "loss/hidden": 3.280859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19141522590070964,
+      "step": 24120
+    },
+    {
+      "epoch": 0.60325,
+      "grad_norm": 28.75,
+      "grad_norm_var": 5.28125,
+      "learning_rate": 0.0001,
+      "loss": 7.248,
+      "loss/crossentropy": 1.9168051473796368,
+      "loss/hidden": 3.31875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17059860788285733,
+      "step": 24130
+    },
+    {
+      "epoch": 0.6035,
+      "grad_norm": 28.25,
+      "grad_norm_var": 4.403059895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.432,
+      "loss/crossentropy": 1.9579944089055061,
+      "loss/hidden": 3.42421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20447726584970952,
+      "step": 24140
+    },
+    {
+      "epoch": 0.60375,
+      "grad_norm": 29.125,
+      "grad_norm_var": 5.737955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4361,
+      "loss/crossentropy": 2.150589424371719,
+      "loss/hidden": 3.359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20431780871003866,
+      "step": 24150
+    },
+    {
+      "epoch": 0.604,
+      "grad_norm": 29.875,
+      "grad_norm_var": 8.734309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3029,
+      "loss/crossentropy": 2.104359980672598,
+      "loss/hidden": 3.314453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19141491586342455,
+      "step": 24160
+    },
+    {
+      "epoch": 0.60425,
+      "grad_norm": 32.25,
+      "grad_norm_var": 4.3416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3517,
+      "loss/crossentropy": 2.0446281746029853,
+      "loss/hidden": 3.35078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18326206738129258,
+      "step": 24170
+    },
+    {
+      "epoch": 0.6045,
+      "grad_norm": 34.5,
+      "grad_norm_var": 35.74791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1912,
+      "loss/crossentropy": 1.9954920992255212,
+      "loss/hidden": 3.32578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17927296683192254,
+      "step": 24180
+    },
+    {
+      "epoch": 0.60475,
+      "grad_norm": 326.0,
+      "grad_norm_var": 5441.927018229167,
+      "learning_rate": 0.0001,
+      "loss": 7.3115,
+      "loss/crossentropy": 1.9973294004797935,
+      "loss/hidden": 3.415625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20168818794190885,
+      "step": 24190
+    },
+    {
+      "epoch": 0.605,
+      "grad_norm": 30.25,
+      "grad_norm_var": 5427.784309895834,
+      "learning_rate": 0.0001,
+      "loss": 7.3283,
+      "loss/crossentropy": 2.075428619980812,
+      "loss/hidden": 3.2828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1890948511660099,
+      "step": 24200
+    },
+    {
+      "epoch": 0.60525,
+      "grad_norm": 37.5,
+      "grad_norm_var": 18.364583333333332,
+      "learning_rate": 0.0001,
+      "loss": 7.3213,
+      "loss/crossentropy": 2.072233888506889,
+      "loss/hidden": 3.25,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17045267168432474,
+      "step": 24210
+    },
+    {
+      "epoch": 0.6055,
+      "grad_norm": 28.5,
+      "grad_norm_var": 6.415559895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3117,
+      "loss/crossentropy": 2.1279248766601087,
+      "loss/hidden": 3.287109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17065610643476248,
+      "step": 24220
+    },
+    {
+      "epoch": 0.60575,
+      "grad_norm": 27.375,
+      "grad_norm_var": 1.4393229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.1783,
+      "loss/crossentropy": 2.0654909342527388,
+      "loss/hidden": 3.29921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1788815937936306,
+      "step": 24230
+    },
+    {
+      "epoch": 0.606,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.3931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2818,
+      "loss/crossentropy": 1.8843905299901962,
+      "loss/hidden": 3.484765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20363699235022067,
+      "step": 24240
+    },
+    {
+      "epoch": 0.60625,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.3238932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2077,
+      "loss/crossentropy": 2.0136945739388468,
+      "loss/hidden": 3.2890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.169308259524405,
+      "step": 24250
+    },
+    {
+      "epoch": 0.6065,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.241080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3067,
+      "loss/crossentropy": 2.06818515509367,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18866120781749487,
+      "step": 24260
+    },
+    {
+      "epoch": 0.60675,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.923958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.35,
+      "loss/crossentropy": 2.1705613493919373,
+      "loss/hidden": 3.328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18546785078942776,
+      "step": 24270
+    },
+    {
+      "epoch": 0.607,
+      "grad_norm": 29.75,
+      "grad_norm_var": 17.083333333333332,
+      "learning_rate": 0.0001,
+      "loss": 7.2704,
+      "loss/crossentropy": 2.1217471458017827,
+      "loss/hidden": 3.42890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20033492222428323,
+      "step": 24280
+    },
+    {
+      "epoch": 0.60725,
+      "grad_norm": 32.5,
+      "grad_norm_var": 14.017643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3065,
+      "loss/crossentropy": 2.013144116103649,
+      "loss/hidden": 3.31015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18037521606311202,
+      "step": 24290
+    },
+    {
+      "epoch": 0.6075,
+      "grad_norm": 31.5,
+      "grad_norm_var": 41.54765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3413,
+      "loss/crossentropy": 1.9536602176725864,
+      "loss/hidden": 3.42265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1871596833691001,
+      "step": 24300
+    },
+    {
+      "epoch": 0.60775,
+      "grad_norm": 29.25,
+      "grad_norm_var": 45.2791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3009,
+      "loss/crossentropy": 2.122793934494257,
+      "loss/hidden": 3.352734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18271594103425742,
+      "step": 24310
+    },
+    {
+      "epoch": 0.608,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.7455729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3036,
+      "loss/crossentropy": 1.9508492700755595,
+      "loss/hidden": 3.423046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1780675446614623,
+      "step": 24320
+    },
+    {
+      "epoch": 0.60825,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.0426432291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3479,
+      "loss/crossentropy": 2.091268266737461,
+      "loss/hidden": 3.303515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1819396572187543,
+      "step": 24330
+    },
+    {
+      "epoch": 0.6085,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.6900390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3685,
+      "loss/crossentropy": 2.1921914756298064,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19144053254276513,
+      "step": 24340
+    },
+    {
+      "epoch": 0.60875,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3547.4488932291665,
+      "learning_rate": 0.0001,
+      "loss": 7.32,
+      "loss/crossentropy": 2.0454428791999817,
+      "loss/hidden": 3.33984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19868660382926465,
+      "step": 24350
+    },
+    {
+      "epoch": 0.609,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.9535807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2958,
+      "loss/crossentropy": 1.978618311882019,
+      "loss/hidden": 3.36484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1801746640354395,
+      "step": 24360
+    },
+    {
+      "epoch": 0.60925,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.381705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2396,
+      "loss/crossentropy": 2.011453700065613,
+      "loss/hidden": 3.285546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17533374354243278,
+      "step": 24370
+    },
+    {
+      "epoch": 0.6095,
+      "grad_norm": 34.5,
+      "grad_norm_var": 4.911393229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2557,
+      "loss/crossentropy": 2.1272184163331986,
+      "loss/hidden": 3.395703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19443935044109822,
+      "step": 24380
+    },
+    {
+      "epoch": 0.60975,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.5712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3372,
+      "loss/crossentropy": 2.0613644808530807,
+      "loss/hidden": 3.41953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1952214989811182,
+      "step": 24390
+    },
+    {
+      "epoch": 0.61,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.2291666666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3664,
+      "loss/crossentropy": 2.082505702972412,
+      "loss/hidden": 3.404296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1836231516674161,
+      "step": 24400
+    },
+    {
+      "epoch": 0.61025,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.94140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4019,
+      "loss/crossentropy": 1.964651158452034,
+      "loss/hidden": 3.501953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18479474298655987,
+      "step": 24410
+    },
+    {
+      "epoch": 0.6105,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.1181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3918,
+      "loss/crossentropy": 2.0536778286099433,
+      "loss/hidden": 3.47734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2254286851733923,
+      "step": 24420
+    },
+    {
+      "epoch": 0.61075,
+      "grad_norm": 28.625,
+      "grad_norm_var": 1.8384765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2519,
+      "loss/crossentropy": 2.0931551665067674,
+      "loss/hidden": 3.461328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19350009728223086,
+      "step": 24430
+    },
+    {
+      "epoch": 0.611,
+      "grad_norm": 28.625,
+      "grad_norm_var": 2.3259765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2784,
+      "loss/crossentropy": 2.1479686103761195,
+      "loss/hidden": 3.271875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17983189076185227,
+      "step": 24440
+    },
+    {
+      "epoch": 0.61125,
+      "grad_norm": 32.75,
+      "grad_norm_var": 7.439322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4267,
+      "loss/crossentropy": 2.171735206246376,
+      "loss/hidden": 3.4703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20030937865376472,
+      "step": 24450
+    },
+    {
+      "epoch": 0.6115,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.9629557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2518,
+      "loss/crossentropy": 1.9000327825546264,
+      "loss/hidden": 3.44375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1778967533260584,
+      "step": 24460
+    },
+    {
+      "epoch": 0.61175,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.87265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3163,
+      "loss/crossentropy": 2.1398353300988675,
+      "loss/hidden": 3.353515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1760717548429966,
+      "step": 24470
+    },
+    {
+      "epoch": 0.612,
+      "grad_norm": 32.25,
+      "grad_norm_var": 4.654166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3402,
+      "loss/crossentropy": 2.0892323054373265,
+      "loss/hidden": 3.377734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18320874478667976,
+      "step": 24480
+    },
+    {
+      "epoch": 0.61225,
+      "grad_norm": 28.25,
+      "grad_norm_var": 26.684375,
+      "learning_rate": 0.0001,
+      "loss": 7.3876,
+      "loss/crossentropy": 1.9319149151444435,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18836453650146723,
+      "step": 24490
+    },
+    {
+      "epoch": 0.6125,
+      "grad_norm": 28.625,
+      "grad_norm_var": 33.487239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3939,
+      "loss/crossentropy": 2.1875485301017763,
+      "loss/hidden": 3.393359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17974788155406712,
+      "step": 24500
+    },
+    {
+      "epoch": 0.61275,
+      "grad_norm": 28.0,
+      "grad_norm_var": 16.7,
+      "learning_rate": 0.0001,
+      "loss": 7.2364,
+      "loss/crossentropy": 2.050992714613676,
+      "loss/hidden": 3.415625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1952893292531371,
+      "step": 24510
+    },
+    {
+      "epoch": 0.613,
+      "grad_norm": 30.75,
+      "grad_norm_var": 9.8212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3393,
+      "loss/crossentropy": 2.1732106573879717,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1879525838419795,
+      "step": 24520
+    },
+    {
+      "epoch": 0.61325,
+      "grad_norm": 31.375,
+      "grad_norm_var": 18.480989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3735,
+      "loss/crossentropy": 2.038769894838333,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18006683513522148,
+      "step": 24530
+    },
+    {
+      "epoch": 0.6135,
+      "grad_norm": 52.25,
+      "grad_norm_var": 3.436879802679899e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2147,
+      "loss/crossentropy": 2.0876556798815726,
+      "loss/hidden": 3.469140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19918184336274863,
+      "step": 24540
+    },
+    {
+      "epoch": 0.61375,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.4368798033364823e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3947,
+      "loss/crossentropy": 2.115431872010231,
+      "loss/hidden": 3.404296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19660551697015763,
+      "step": 24550
+    },
+    {
+      "epoch": 0.614,
+      "grad_norm": 29.875,
+      "grad_norm_var": 10.852018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.1815,
+      "loss/crossentropy": 2.01566364467144,
+      "loss/hidden": 3.36015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1908690959215164,
+      "step": 24560
+    },
+    {
+      "epoch": 0.61425,
+      "grad_norm": 29.25,
+      "grad_norm_var": 10.713541666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2148,
+      "loss/crossentropy": 1.9971563063561917,
+      "loss/hidden": 3.437890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19316924586892129,
+      "step": 24570
+    },
+    {
+      "epoch": 0.6145,
+      "grad_norm": 28.75,
+      "grad_norm_var": 5.447330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.295,
+      "loss/crossentropy": 2.1087226420640945,
+      "loss/hidden": 3.42578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1953347373753786,
+      "step": 24580
+    },
+    {
+      "epoch": 0.61475,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.9184895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2813,
+      "loss/crossentropy": 2.0959045618772505,
+      "loss/hidden": 3.362890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.183117344789207,
+      "step": 24590
+    },
+    {
+      "epoch": 0.615,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.040559895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2361,
+      "loss/crossentropy": 1.9878082022070884,
+      "loss/hidden": 3.31875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17381655257195233,
+      "step": 24600
+    },
+    {
+      "epoch": 0.61525,
+      "grad_norm": 30.875,
+      "grad_norm_var": 5.760416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3343,
+      "loss/crossentropy": 1.9321263507008553,
+      "loss/hidden": 3.39296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1903281847015023,
+      "step": 24610
+    },
+    {
+      "epoch": 0.6155,
+      "grad_norm": 31.25,
+      "grad_norm_var": 5.926822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2588,
+      "loss/crossentropy": 2.2268902391195295,
+      "loss/hidden": 3.31953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19283902421593666,
+      "step": 24620
+    },
+    {
+      "epoch": 0.61575,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.1416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3226,
+      "loss/crossentropy": 2.0439958930015565,
+      "loss/hidden": 3.393359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18890973944216966,
+      "step": 24630
+    },
+    {
+      "epoch": 0.616,
+      "grad_norm": 34.0,
+      "grad_norm_var": 2.653059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2599,
+      "loss/crossentropy": 1.945370616018772,
+      "loss/hidden": 3.35625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17533454354852437,
+      "step": 24640
+    },
+    {
+      "epoch": 0.61625,
+      "grad_norm": 29.625,
+      "grad_norm_var": 6.738541666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2148,
+      "loss/crossentropy": 1.991739486157894,
+      "loss/hidden": 3.35390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18211280852556228,
+      "step": 24650
+    },
+    {
+      "epoch": 0.6165,
+      "grad_norm": 29.25,
+      "grad_norm_var": 14.560416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3611,
+      "loss/crossentropy": 2.075421932339668,
+      "loss/hidden": 3.453515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19871473647654056,
+      "step": 24660
+    },
+    {
+      "epoch": 0.61675,
+      "grad_norm": 29.75,
+      "grad_norm_var": 12.080989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.239,
+      "loss/crossentropy": 2.008017046749592,
+      "loss/hidden": 3.27265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17111360877752305,
+      "step": 24670
+    },
+    {
+      "epoch": 0.617,
+      "grad_norm": 28.25,
+      "grad_norm_var": 8.795833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2497,
+      "loss/crossentropy": 1.9377702206373215,
+      "loss/hidden": 3.349609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17999718375504017,
+      "step": 24680
+    },
+    {
+      "epoch": 0.61725,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.385872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3279,
+      "loss/crossentropy": 2.1406032904982566,
+      "loss/hidden": 3.281640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1813998430967331,
+      "step": 24690
+    },
+    {
+      "epoch": 0.6175,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.7212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3508,
+      "loss/crossentropy": 2.074493780732155,
+      "loss/hidden": 3.39375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1942134676501155,
+      "step": 24700
+    },
+    {
+      "epoch": 0.61775,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.9275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3579,
+      "loss/crossentropy": 2.235245580971241,
+      "loss/hidden": 3.337109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1957023985683918,
+      "step": 24710
+    },
+    {
+      "epoch": 0.618,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.6817057291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3896,
+      "loss/crossentropy": 2.0386796444654465,
+      "loss/hidden": 3.326953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19206191059201955,
+      "step": 24720
+    },
+    {
+      "epoch": 0.61825,
+      "grad_norm": 29.625,
+      "grad_norm_var": 1.7705729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3513,
+      "loss/crossentropy": 2.0323100224137307,
+      "loss/hidden": 3.4015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20894596744328736,
+      "step": 24730
+    },
+    {
+      "epoch": 0.6185,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.4322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2488,
+      "loss/crossentropy": 2.0485980585217476,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18473935313522816,
+      "step": 24740
+    },
+    {
+      "epoch": 0.61875,
+      "grad_norm": 34.0,
+      "grad_norm_var": 1.8436848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3125,
+      "loss/crossentropy": 2.12614818662405,
+      "loss/hidden": 3.460546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18453964591026306,
+      "step": 24750
+    },
+    {
+      "epoch": 0.619,
+      "grad_norm": 35.25,
+      "grad_norm_var": 3.7080729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.307,
+      "loss/crossentropy": 2.0480948239564896,
+      "loss/hidden": 3.367578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18478762162849308,
+      "step": 24760
+    },
+    {
+      "epoch": 0.61925,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.8223307291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2388,
+      "loss/crossentropy": 2.007586943730712,
+      "loss/hidden": 3.32421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1709001273382455,
+      "step": 24770
+    },
+    {
+      "epoch": 0.6195,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.1497395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2203,
+      "loss/crossentropy": 1.9536567956209183,
+      "loss/hidden": 3.423046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18609143160283564,
+      "step": 24780
+    },
+    {
+      "epoch": 0.61975,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.158333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4193,
+      "loss/crossentropy": 2.1475788801908493,
+      "loss/hidden": 3.37890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19103976357728242,
+      "step": 24790
+    },
+    {
+      "epoch": 0.62,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.4150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2597,
+      "loss/crossentropy": 1.93576335683465,
+      "loss/hidden": 3.26015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1649831486865878,
+      "step": 24800
+    },
+    {
+      "epoch": 0.62025,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.064322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.264,
+      "loss/crossentropy": 2.000751170516014,
+      "loss/hidden": 3.35,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17869363045319914,
+      "step": 24810
+    },
+    {
+      "epoch": 0.6205,
+      "grad_norm": 31.5,
+      "grad_norm_var": 4.509830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3578,
+      "loss/crossentropy": 2.100572241842747,
+      "loss/hidden": 3.370703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19480398446321487,
+      "step": 24820
+    },
+    {
+      "epoch": 0.62075,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.252795170612926e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3303,
+      "loss/crossentropy": 2.0044652953743936,
+      "loss/hidden": 3.444921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19480419661849738,
+      "step": 24830
+    },
+    {
+      "epoch": 0.621,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.0885416666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3171,
+      "loss/crossentropy": 2.163808681815863,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18681660266593098,
+      "step": 24840
+    },
+    {
+      "epoch": 0.62125,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.4400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.333,
+      "loss/crossentropy": 2.03867654427886,
+      "loss/hidden": 3.288671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1776962866075337,
+      "step": 24850
+    },
+    {
+      "epoch": 0.6215,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.17265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3692,
+      "loss/crossentropy": 2.0644729137420654,
+      "loss/hidden": 3.509375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1982646532356739,
+      "step": 24860
+    },
+    {
+      "epoch": 0.62175,
+      "grad_norm": 27.625,
+      "grad_norm_var": 3.06015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3453,
+      "loss/crossentropy": 2.171038728952408,
+      "loss/hidden": 3.31328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19105625264346598,
+      "step": 24870
+    },
+    {
+      "epoch": 0.622,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.09140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3096,
+      "loss/crossentropy": 2.0491355784237384,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19366521965712308,
+      "step": 24880
+    },
+    {
+      "epoch": 0.62225,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.001822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2684,
+      "loss/crossentropy": 2.185651781409979,
+      "loss/hidden": 3.2890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1791619287803769,
+      "step": 24890
+    },
+    {
+      "epoch": 0.6225,
+      "grad_norm": 28.25,
+      "grad_norm_var": 2.0056640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2828,
+      "loss/crossentropy": 2.1073651291429996,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19452255237847565,
+      "step": 24900
+    },
+    {
+      "epoch": 0.62275,
+      "grad_norm": 28.25,
+      "grad_norm_var": 2.814583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.324,
+      "loss/crossentropy": 2.1819552317261697,
+      "loss/hidden": 3.47734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19712742660194635,
+      "step": 24910
+    },
+    {
+      "epoch": 0.623,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.340625,
+      "learning_rate": 0.0001,
+      "loss": 7.2135,
+      "loss/crossentropy": 2.0284355215728285,
+      "loss/hidden": 3.304296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17623181734234095,
+      "step": 24920
+    },
+    {
+      "epoch": 0.62325,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.7405598958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3556,
+      "loss/crossentropy": 1.9424869917333125,
+      "loss/hidden": 3.334375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16103002466261387,
+      "step": 24930
+    },
+    {
+      "epoch": 0.6235,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.687239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4098,
+      "loss/crossentropy": 2.1412764593958853,
+      "loss/hidden": 3.426953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18787063639611007,
+      "step": 24940
+    },
+    {
+      "epoch": 0.62375,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.4374348958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.423,
+      "loss/crossentropy": 2.1100604355335237,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1824411964043975,
+      "step": 24950
+    },
+    {
+      "epoch": 0.624,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.1405598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2604,
+      "loss/crossentropy": 2.173184335231781,
+      "loss/hidden": 3.3046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18387555023655294,
+      "step": 24960
+    },
+    {
+      "epoch": 0.62425,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.775,
+      "learning_rate": 0.0001,
+      "loss": 7.4108,
+      "loss/crossentropy": 2.2970198541879654,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1978620221838355,
+      "step": 24970
+    },
+    {
+      "epoch": 0.6245,
+      "grad_norm": 30.25,
+      "grad_norm_var": 6.2869140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3759,
+      "loss/crossentropy": 2.098717086017132,
+      "loss/hidden": 3.2953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1739401025697589,
+      "step": 24980
+    },
+    {
+      "epoch": 0.62475,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.34765625,
+      "learning_rate": 0.0001,
+      "loss": 7.1911,
+      "loss/crossentropy": 2.0788547694683075,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18489116076380013,
+      "step": 24990
+    },
+    {
+      "epoch": 0.625,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.3181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.289,
+      "loss/crossentropy": 2.05898377597332,
+      "loss/hidden": 3.396484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19787134062498807,
+      "step": 25000
+    },
+    {
+      "epoch": 0.62525,
+      "grad_norm": 30.5,
+      "grad_norm_var": 5.060416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3426,
+      "loss/crossentropy": 2.191389924287796,
+      "loss/hidden": 3.35234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1947077002376318,
+      "step": 25010
+    },
+    {
+      "epoch": 0.6255,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.7660807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3552,
+      "loss/crossentropy": 2.1716764003038405,
+      "loss/hidden": 3.39296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18750451393425466,
+      "step": 25020
+    },
+    {
+      "epoch": 0.62575,
+      "grad_norm": 28.75,
+      "grad_norm_var": 3.3934895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2251,
+      "loss/crossentropy": 2.0290740236639975,
+      "loss/hidden": 3.44296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2046494733542204,
+      "step": 25030
+    },
+    {
+      "epoch": 0.626,
+      "grad_norm": 29.125,
+      "grad_norm_var": 15.915559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3018,
+      "loss/crossentropy": 2.0449746288359165,
+      "loss/hidden": 3.36328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17823896156623958,
+      "step": 25040
+    },
+    {
+      "epoch": 0.62625,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.39140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2551,
+      "loss/crossentropy": 2.0032063253223895,
+      "loss/hidden": 3.443359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18292871192097665,
+      "step": 25050
+    },
+    {
+      "epoch": 0.6265,
+      "grad_norm": 34.25,
+      "grad_norm_var": 2.574934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3934,
+      "loss/crossentropy": 1.9866469122469426,
+      "loss/hidden": 3.444921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18698836211115122,
+      "step": 25060
+    },
+    {
+      "epoch": 0.62675,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.9202473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3007,
+      "loss/crossentropy": 2.0594868302345275,
+      "loss/hidden": 3.386328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18997351210564375,
+      "step": 25070
+    },
+    {
+      "epoch": 0.627,
+      "grad_norm": 34.5,
+      "grad_norm_var": 3.217643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2211,
+      "loss/crossentropy": 2.0622723668813707,
+      "loss/hidden": 3.404296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18301580119878053,
+      "step": 25080
+    },
+    {
+      "epoch": 0.62725,
+      "grad_norm": 30.875,
+      "grad_norm_var": 6.513541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2187,
+      "loss/crossentropy": 2.19646572843194,
+      "loss/hidden": 3.41015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21105999127030373,
+      "step": 25090
+    },
+    {
+      "epoch": 0.6275,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.287239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3144,
+      "loss/crossentropy": 2.1136497229337694,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18730370607227087,
+      "step": 25100
+    },
+    {
+      "epoch": 0.62775,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.6087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4126,
+      "loss/crossentropy": 2.2990898922085763,
+      "loss/hidden": 3.38359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.192084638774395,
+      "step": 25110
+    },
+    {
+      "epoch": 0.628,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.9275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3276,
+      "loss/crossentropy": 1.8526959542185069,
+      "loss/hidden": 3.478125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19008365646004677,
+      "step": 25120
+    },
+    {
+      "epoch": 0.62825,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.4181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.1824,
+      "loss/crossentropy": 1.9762011893093585,
+      "loss/hidden": 3.33203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.170592157356441,
+      "step": 25130
+    },
+    {
+      "epoch": 0.6285,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.67890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3663,
+      "loss/crossentropy": 2.0375132709741592,
+      "loss/hidden": 3.42265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19559677615761756,
+      "step": 25140
+    },
+    {
+      "epoch": 0.62875,
+      "grad_norm": 5268045824.0,
+      "grad_norm_var": 1.7345191550355768e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2744,
+      "loss/crossentropy": 2.116751839220524,
+      "loss/hidden": 3.64453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20043763648718596,
+      "step": 25150
+    },
+    {
+      "epoch": 0.629,
+      "grad_norm": 29.625,
+      "grad_norm_var": 1.7345191549916764e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.1578,
+      "loss/crossentropy": 1.8911869063973428,
+      "loss/hidden": 3.3375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17289240173995496,
+      "step": 25160
+    },
+    {
+      "epoch": 0.62925,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.628125,
+      "learning_rate": 0.0001,
+      "loss": 7.2708,
+      "loss/crossentropy": 1.9820346303284169,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17856419319286942,
+      "step": 25170
+    },
+    {
+      "epoch": 0.6295,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.124739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2472,
+      "loss/crossentropy": 2.166484069824219,
+      "loss/hidden": 3.361328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1979134999215603,
+      "step": 25180
+    },
+    {
+      "epoch": 0.62975,
+      "grad_norm": 29.875,
+      "grad_norm_var": 15.1025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2973,
+      "loss/crossentropy": 2.0940047293901443,
+      "loss/hidden": 3.4125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19144841339439153,
+      "step": 25190
+    },
+    {
+      "epoch": 0.63,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.3934895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2884,
+      "loss/crossentropy": 2.0762594789266586,
+      "loss/hidden": 3.377734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1935945412144065,
+      "step": 25200
+    },
+    {
+      "epoch": 0.63025,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.083268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3037,
+      "loss/crossentropy": 2.1710478901863097,
+      "loss/hidden": 3.333984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19344222452491522,
+      "step": 25210
+    },
+    {
+      "epoch": 0.6305,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.512434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3412,
+      "loss/crossentropy": 2.1715824723243715,
+      "loss/hidden": 3.367578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18924503941088916,
+      "step": 25220
+    },
+    {
+      "epoch": 0.63075,
+      "grad_norm": 28.125,
+      "grad_norm_var": 3.0830729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2856,
+      "loss/crossentropy": 2.1482178628444673,
+      "loss/hidden": 3.275,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1806714640930295,
+      "step": 25230
+    },
+    {
+      "epoch": 0.631,
+      "grad_norm": 31.25,
+      "grad_norm_var": 4.3875,
+      "learning_rate": 0.0001,
+      "loss": 7.3432,
+      "loss/crossentropy": 1.8948672696948052,
+      "loss/hidden": 3.459375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18248430043458938,
+      "step": 25240
+    },
+    {
+      "epoch": 0.63125,
+      "grad_norm": 30.625,
+      "grad_norm_var": 5.961393229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.307,
+      "loss/crossentropy": 1.979700568318367,
+      "loss/hidden": 3.368359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18111934289336204,
+      "step": 25250
+    },
+    {
+      "epoch": 0.6315,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.0884765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3091,
+      "loss/crossentropy": 2.1454955518245695,
+      "loss/hidden": 3.433203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20133222881704568,
+      "step": 25260
+    },
+    {
+      "epoch": 0.63175,
+      "grad_norm": 32.25,
+      "grad_norm_var": 11.745247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3298,
+      "loss/crossentropy": 2.052339327335358,
+      "loss/hidden": 3.368359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1785123085603118,
+      "step": 25270
+    },
+    {
+      "epoch": 0.632,
+      "grad_norm": 30.25,
+      "grad_norm_var": 16.828580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4661,
+      "loss/crossentropy": 2.0869883239269256,
+      "loss/hidden": 3.40859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2130020596086979,
+      "step": 25280
+    },
+    {
+      "epoch": 0.63225,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.2916015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2918,
+      "loss/crossentropy": 1.994824843108654,
+      "loss/hidden": 3.207421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17445887122303247,
+      "step": 25290
+    },
+    {
+      "epoch": 0.6325,
+      "grad_norm": 35.0,
+      "grad_norm_var": 116.28098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2917,
+      "loss/crossentropy": 2.073824466764927,
+      "loss/hidden": 3.358203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18123935330659152,
+      "step": 25300
+    },
+    {
+      "epoch": 0.63275,
+      "grad_norm": 30.625,
+      "grad_norm_var": 195.25774739583332,
+      "learning_rate": 0.0001,
+      "loss": 7.293,
+      "loss/crossentropy": 2.102488835155964,
+      "loss/hidden": 3.362890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1843341593630612,
+      "step": 25310
+    },
+    {
+      "epoch": 0.633,
+      "grad_norm": 30.125,
+      "grad_norm_var": 103.74973958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2951,
+      "loss/crossentropy": 1.9738748401403428,
+      "loss/hidden": 3.38671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19301892705261708,
+      "step": 25320
+    },
+    {
+      "epoch": 0.63325,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.2635411529232722e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3154,
+      "loss/crossentropy": 2.1203658998012545,
+      "loss/hidden": 3.274609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1767484113574028,
+      "step": 25330
+    },
+    {
+      "epoch": 0.6335,
+      "grad_norm": 28.0,
+      "grad_norm_var": 1.2635411512184276e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3058,
+      "loss/crossentropy": 2.0557747781276703,
+      "loss/hidden": 3.344140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18306177966296672,
+      "step": 25340
+    },
+    {
+      "epoch": 0.63375,
+      "grad_norm": 28.625,
+      "grad_norm_var": 34.753125,
+      "learning_rate": 0.0001,
+      "loss": 7.0454,
+      "loss/crossentropy": 2.0685575902462006,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1882263720035553,
+      "step": 25350
+    },
+    {
+      "epoch": 0.634,
+      "grad_norm": 28.875,
+      "grad_norm_var": 22.106705729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2534,
+      "loss/crossentropy": 2.208861267566681,
+      "loss/hidden": 3.339453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.197300866805017,
+      "step": 25360
+    },
+    {
+      "epoch": 0.63425,
+      "grad_norm": 28.125,
+      "grad_norm_var": 10.518489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1867,
+      "loss/crossentropy": 2.133610498905182,
+      "loss/hidden": 3.48203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1941371016204357,
+      "step": 25370
+    },
+    {
+      "epoch": 0.6345,
+      "grad_norm": 28.875,
+      "grad_norm_var": 21.129166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2715,
+      "loss/crossentropy": 2.063799539208412,
+      "loss/hidden": 3.273828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1706739742308855,
+      "step": 25380
+    },
+    {
+      "epoch": 0.63475,
+      "grad_norm": 38.25,
+      "grad_norm_var": 14.7416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.0581,
+      "loss/crossentropy": 2.0165546610951424,
+      "loss/hidden": 3.141796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.15498318960890173,
+      "step": 25390
+    },
+    {
+      "epoch": 0.635,
+      "grad_norm": 37.75,
+      "grad_norm_var": 14.0353515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2961,
+      "loss/crossentropy": 2.240743267536163,
+      "loss/hidden": 3.30546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18806628789752722,
+      "step": 25400
+    },
+    {
+      "epoch": 0.63525,
+      "grad_norm": 33.25,
+      "grad_norm_var": 19.630989583333335,
+      "learning_rate": 0.0001,
+      "loss": 7.1944,
+      "loss/crossentropy": 2.1175243996083735,
+      "loss/hidden": 3.28046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18017748612910509,
+      "step": 25410
+    },
+    {
+      "epoch": 0.6355,
+      "grad_norm": 30.375,
+      "grad_norm_var": 16.4025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2159,
+      "loss/crossentropy": 2.11346255838871,
+      "loss/hidden": 3.270703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1687102049589157,
+      "step": 25420
+    },
+    {
+      "epoch": 0.63575,
+      "grad_norm": 31.75,
+      "grad_norm_var": 41.88014322916667,
+      "learning_rate": 0.0001,
+      "loss": 7.2323,
+      "loss/crossentropy": 2.2356766797602177,
+      "loss/hidden": 3.31484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18139579724520444,
+      "step": 25430
+    },
+    {
+      "epoch": 0.636,
+      "grad_norm": 30.75,
+      "grad_norm_var": 13.415559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1997,
+      "loss/crossentropy": 1.9421189457178116,
+      "loss/hidden": 3.35625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17150872349739074,
+      "step": 25440
+    },
+    {
+      "epoch": 0.63625,
+      "grad_norm": 29.625,
+      "grad_norm_var": 5.9775390625,
+      "learning_rate": 0.0001,
+      "loss": 7.1837,
+      "loss/crossentropy": 2.0306406378746034,
+      "loss/hidden": 3.2375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17749902671203016,
+      "step": 25450
+    },
+    {
+      "epoch": 0.6365,
+      "grad_norm": 30.0,
+      "grad_norm_var": 8156.041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2467,
+      "loss/crossentropy": 2.0117203891277313,
+      "loss/hidden": 3.301953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17507544625550508,
+      "step": 25460
+    },
+    {
+      "epoch": 0.63675,
+      "grad_norm": 36.75,
+      "grad_norm_var": 18.21015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2268,
+      "loss/crossentropy": 2.0377916976809503,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17452282812446357,
+      "step": 25470
+    },
+    {
+      "epoch": 0.637,
+      "grad_norm": 29.625,
+      "grad_norm_var": 26.654166666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2038,
+      "loss/crossentropy": 2.0024905622005464,
+      "loss/hidden": 3.4453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18442925848066807,
+      "step": 25480
+    },
+    {
+      "epoch": 0.63725,
+      "grad_norm": 27.125,
+      "grad_norm_var": 28.959309895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.2125,
+      "loss/crossentropy": 2.0766751676797868,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20771742928773165,
+      "step": 25490
+    },
+    {
+      "epoch": 0.6375,
+      "grad_norm": 34.75,
+      "grad_norm_var": 6.7650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.1703,
+      "loss/crossentropy": 2.0064065635204313,
+      "loss/hidden": 3.289453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18250296991318465,
+      "step": 25500
+    },
+    {
+      "epoch": 0.63775,
+      "grad_norm": 30.75,
+      "grad_norm_var": 4.0697265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2983,
+      "loss/crossentropy": 2.0614541321992874,
+      "loss/hidden": 3.442578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2076763341203332,
+      "step": 25510
+    },
+    {
+      "epoch": 0.638,
+      "grad_norm": 29.5,
+      "grad_norm_var": 7.487239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2289,
+      "loss/crossentropy": 1.9859767436981202,
+      "loss/hidden": 3.2984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17486365381628274,
+      "step": 25520
+    },
+    {
+      "epoch": 0.63825,
+      "grad_norm": 35.75,
+      "grad_norm_var": 17.522916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2378,
+      "loss/crossentropy": 1.9697401717305183,
+      "loss/hidden": 3.40625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1799583438783884,
+      "step": 25530
+    },
+    {
+      "epoch": 0.6385,
+      "grad_norm": 28.375,
+      "grad_norm_var": 38.739322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2664,
+      "loss/crossentropy": 2.1364367194473743,
+      "loss/hidden": 3.290625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18297823830507695,
+      "step": 25540
+    },
+    {
+      "epoch": 0.63875,
+      "grad_norm": 30.125,
+      "grad_norm_var": 35.773372395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.1437,
+      "loss/crossentropy": 2.0668285630643366,
+      "loss/hidden": 3.330078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1804805461317301,
+      "step": 25550
+    },
+    {
+      "epoch": 0.639,
+      "grad_norm": 28.625,
+      "grad_norm_var": 11.2619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2275,
+      "loss/crossentropy": 2.093304879963398,
+      "loss/hidden": 3.273046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18014004416763782,
+      "step": 25560
+    },
+    {
+      "epoch": 0.63925,
+      "grad_norm": 35.0,
+      "grad_norm_var": 10.3150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3097,
+      "loss/crossentropy": 2.20695668309927,
+      "loss/hidden": 3.383984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19993221275508405,
+      "step": 25570
+    },
+    {
+      "epoch": 0.6395,
+      "grad_norm": 35.5,
+      "grad_norm_var": 6.303125,
+      "learning_rate": 0.0001,
+      "loss": 7.257,
+      "loss/crossentropy": 2.0658087216317655,
+      "loss/hidden": 3.303125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18781778533011675,
+      "step": 25580
+    },
+    {
+      "epoch": 0.63975,
+      "grad_norm": 29.875,
+      "grad_norm_var": 4.42265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2849,
+      "loss/crossentropy": 2.1931436568498612,
+      "loss/hidden": 3.33515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18365668430924414,
+      "step": 25590
+    },
+    {
+      "epoch": 0.64,
+      "grad_norm": 28.125,
+      "grad_norm_var": 4.941666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3091,
+      "loss/crossentropy": 2.105237030982971,
+      "loss/hidden": 3.301953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18291805479675532,
+      "step": 25600
+    },
+    {
+      "epoch": 0.64025,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.7983723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3204,
+      "loss/crossentropy": 1.995211163163185,
+      "loss/hidden": 3.383984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17982382252812384,
+      "step": 25610
+    },
+    {
+      "epoch": 0.6405,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.4504557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4159,
+      "loss/crossentropy": 2.0612260174006223,
+      "loss/hidden": 3.3734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19814395047724248,
+      "step": 25620
+    },
+    {
+      "epoch": 0.64075,
+      "grad_norm": 29.25,
+      "grad_norm_var": 1.5655598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3244,
+      "loss/crossentropy": 2.1825088411569595,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17284797839820384,
+      "step": 25630
+    },
+    {
+      "epoch": 0.641,
+      "grad_norm": 30.625,
+      "grad_norm_var": 4.470572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3127,
+      "loss/crossentropy": 2.123421373963356,
+      "loss/hidden": 3.304296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1951315425336361,
+      "step": 25640
+    },
+    {
+      "epoch": 0.64125,
+      "grad_norm": 29.25,
+      "grad_norm_var": 2.6211653257674097e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4566,
+      "loss/crossentropy": 1.9836766071617604,
+      "loss/hidden": 3.338671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18384753745049237,
+      "step": 25650
+    },
+    {
+      "epoch": 0.6415,
+      "grad_norm": 30.125,
+      "grad_norm_var": 10.3619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2435,
+      "loss/crossentropy": 2.024305807799101,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18380997078493239,
+      "step": 25660
+    },
+    {
+      "epoch": 0.64175,
+      "grad_norm": 29.25,
+      "grad_norm_var": 6.990559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3449,
+      "loss/crossentropy": 2.0873545318841935,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1851763878017664,
+      "step": 25670
+    },
+    {
+      "epoch": 0.642,
+      "grad_norm": 29.75,
+      "grad_norm_var": 6.716080729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3962,
+      "loss/crossentropy": 1.953444430232048,
+      "loss/hidden": 3.45625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18732787277549506,
+      "step": 25680
+    },
+    {
+      "epoch": 0.64225,
+      "grad_norm": 30.0,
+      "grad_norm_var": 4.412434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3693,
+      "loss/crossentropy": 2.0779720529913903,
+      "loss/hidden": 3.45625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19645520951598883,
+      "step": 25690
+    },
+    {
+      "epoch": 0.6425,
+      "grad_norm": 36.0,
+      "grad_norm_var": 6.120247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2773,
+      "loss/crossentropy": 2.080344262719154,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18719701357185842,
+      "step": 25700
+    },
+    {
+      "epoch": 0.64275,
+      "grad_norm": 31.75,
+      "grad_norm_var": 5.784309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2415,
+      "loss/crossentropy": 2.1494850531220435,
+      "loss/hidden": 3.44375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2048310289159417,
+      "step": 25710
+    },
+    {
+      "epoch": 0.643,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.9572916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1719,
+      "loss/crossentropy": 2.06035181209445,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19010064098984003,
+      "step": 25720
+    },
+    {
+      "epoch": 0.64325,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.4150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3085,
+      "loss/crossentropy": 1.9468969091773034,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18584002749994397,
+      "step": 25730
+    },
+    {
+      "epoch": 0.6435,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.221875,
+      "learning_rate": 0.0001,
+      "loss": 7.2789,
+      "loss/crossentropy": 1.9196617759764194,
+      "loss/hidden": 3.416796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17771600587293507,
+      "step": 25740
+    },
+    {
+      "epoch": 0.64375,
+      "grad_norm": 30.25,
+      "grad_norm_var": 14.9853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2678,
+      "loss/crossentropy": 2.1726732790470122,
+      "loss/hidden": 3.395703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18251333478838205,
+      "step": 25750
+    },
+    {
+      "epoch": 0.644,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.2059895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3832,
+      "loss/crossentropy": 2.1877689242362974,
+      "loss/hidden": 3.36640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19289560578763484,
+      "step": 25760
+    },
+    {
+      "epoch": 0.64425,
+      "grad_norm": 62.75,
+      "grad_norm_var": 69.259375,
+      "learning_rate": 0.0001,
+      "loss": 7.2773,
+      "loss/crossentropy": 2.070755659043789,
+      "loss/hidden": 3.3203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17988401111215352,
+      "step": 25770
+    },
+    {
+      "epoch": 0.6445,
+      "grad_norm": 29.375,
+      "grad_norm_var": 113.71399739583333,
+      "learning_rate": 0.0001,
+      "loss": 7.3734,
+      "loss/crossentropy": 2.106437236815691,
+      "loss/hidden": 3.372265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19078649543225765,
+      "step": 25780
+    },
+    {
+      "epoch": 0.64475,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.3952473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3165,
+      "loss/crossentropy": 2.1690757513046264,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19081941563636065,
+      "step": 25790
+    },
+    {
+      "epoch": 0.645,
+      "grad_norm": 28.75,
+      "grad_norm_var": 1.6077473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3326,
+      "loss/crossentropy": 2.1447702586650848,
+      "loss/hidden": 3.35859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19072114508599042,
+      "step": 25800
+    },
+    {
+      "epoch": 0.64525,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.5410807291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3061,
+      "loss/crossentropy": 2.243308076262474,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19677081871777774,
+      "step": 25810
+    },
+    {
+      "epoch": 0.6455,
+      "grad_norm": 29.25,
+      "grad_norm_var": 2.2811848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2598,
+      "loss/crossentropy": 2.1137851983308793,
+      "loss/hidden": 3.306640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18019928000867366,
+      "step": 25820
+    },
+    {
+      "epoch": 0.64575,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.3067057291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2768,
+      "loss/crossentropy": 2.046476516127586,
+      "loss/hidden": 3.396875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1881034353747964,
+      "step": 25830
+    },
+    {
+      "epoch": 0.646,
+      "grad_norm": 30.875,
+      "grad_norm_var": 7.928059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4478,
+      "loss/crossentropy": 1.9755774058401585,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18677257960662247,
+      "step": 25840
+    },
+    {
+      "epoch": 0.64625,
+      "grad_norm": 29.0,
+      "grad_norm_var": 10.15625,
+      "learning_rate": 0.0001,
+      "loss": 7.35,
+      "loss/crossentropy": 2.0693862795829774,
+      "loss/hidden": 3.3375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18849658444523812,
+      "step": 25850
+    },
+    {
+      "epoch": 0.6465,
+      "grad_norm": 30.125,
+      "grad_norm_var": 10.2125,
+      "learning_rate": 0.0001,
+      "loss": 7.3352,
+      "loss/crossentropy": 2.121247774362564,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20228977780789137,
+      "step": 25860
+    },
+    {
+      "epoch": 0.64675,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.5160807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3408,
+      "loss/crossentropy": 2.0337157592177393,
+      "loss/hidden": 3.45390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18892905581742525,
+      "step": 25870
+    },
+    {
+      "epoch": 0.647,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.974739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3046,
+      "loss/crossentropy": 1.9749686062335967,
+      "loss/hidden": 3.486328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19374344814568759,
+      "step": 25880
+    },
+    {
+      "epoch": 0.64725,
+      "grad_norm": 33.25,
+      "grad_norm_var": 5.467708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3869,
+      "loss/crossentropy": 2.083652201294899,
+      "loss/hidden": 3.392578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2009931992739439,
+      "step": 25890
+    },
+    {
+      "epoch": 0.6475,
+      "grad_norm": 34.5,
+      "grad_norm_var": 4.03515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3113,
+      "loss/crossentropy": 2.0592153131961823,
+      "loss/hidden": 3.332421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19190128184854985,
+      "step": 25900
+    },
+    {
+      "epoch": 0.64775,
+      "grad_norm": 28.625,
+      "grad_norm_var": 4.687955729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2909,
+      "loss/crossentropy": 1.9693334527313708,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1817644027993083,
+      "step": 25910
+    },
+    {
+      "epoch": 0.648,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.124739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3546,
+      "loss/crossentropy": 2.079821580648422,
+      "loss/hidden": 3.3125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17972610034048558,
+      "step": 25920
+    },
+    {
+      "epoch": 0.64825,
+      "grad_norm": 28.375,
+      "grad_norm_var": 3.4934895833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2074,
+      "loss/crossentropy": 2.1161312617361547,
+      "loss/hidden": 3.34296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1969278533011675,
+      "step": 25930
+    },
+    {
+      "epoch": 0.6485,
+      "grad_norm": 27.25,
+      "grad_norm_var": 1.8212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3702,
+      "loss/crossentropy": 2.1124040342867376,
+      "loss/hidden": 3.379296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18695534877479075,
+      "step": 25940
+    },
+    {
+      "epoch": 0.64875,
+      "grad_norm": 34.25,
+      "grad_norm_var": 4.521809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4366,
+      "loss/crossentropy": 2.1193448536098005,
+      "loss/hidden": 3.36484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18631769344210625,
+      "step": 25950
+    },
+    {
+      "epoch": 0.649,
+      "grad_norm": 30.875,
+      "grad_norm_var": 4.4697265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4333,
+      "loss/crossentropy": 2.1864205464720725,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19288004878908396,
+      "step": 25960
+    },
+    {
+      "epoch": 0.64925,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.65390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3549,
+      "loss/crossentropy": 2.0109216958284377,
+      "loss/hidden": 3.345703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18039658814668655,
+      "step": 25970
+    },
+    {
+      "epoch": 0.6495,
+      "grad_norm": 6677331968.0,
+      "grad_norm_var": 2.786672611783344e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3897,
+      "loss/crossentropy": 1.9869451895356178,
+      "loss/hidden": 3.3234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17617017943412067,
+      "step": 25980
+    },
+    {
+      "epoch": 0.64975,
+      "grad_norm": 29.25,
+      "grad_norm_var": 2.7866726110599665e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2854,
+      "loss/crossentropy": 2.2265473932027815,
+      "loss/hidden": 3.3734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2203309465199709,
+      "step": 25990
+    },
+    {
+      "epoch": 0.65,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.5822265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3065,
+      "loss/crossentropy": 2.0998080484569073,
+      "loss/hidden": 3.4484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18724181046709418,
+      "step": 26000
+    },
+    {
+      "epoch": 0.65025,
+      "grad_norm": 29.0,
+      "grad_norm_var": 9.589518229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.0924,
+      "loss/crossentropy": 1.897086238861084,
+      "loss/hidden": 3.268359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16815462270751597,
+      "step": 26010
+    },
+    {
+      "epoch": 0.6505,
+      "grad_norm": 28.75,
+      "grad_norm_var": 9.5869140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2116,
+      "loss/crossentropy": 2.1296266965568065,
+      "loss/hidden": 3.293359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18028478175401688,
+      "step": 26020
+    },
+    {
+      "epoch": 0.65075,
+      "grad_norm": 28.75,
+      "grad_norm_var": 3.5872395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2718,
+      "loss/crossentropy": 2.0453512027859686,
+      "loss/hidden": 3.26484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17293851375579833,
+      "step": 26030
+    },
+    {
+      "epoch": 0.651,
+      "grad_norm": 32.0,
+      "grad_norm_var": 24.023372395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3717,
+      "loss/crossentropy": 2.1030614957213403,
+      "loss/hidden": 3.29296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17731762621551753,
+      "step": 26040
+    },
+    {
+      "epoch": 0.65125,
+      "grad_norm": 28.75,
+      "grad_norm_var": 9.29140625,
+      "learning_rate": 0.0001,
+      "loss": 7.1228,
+      "loss/crossentropy": 2.021975876390934,
+      "loss/hidden": 3.36328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18740149699151515,
+      "step": 26050
+    },
+    {
+      "epoch": 0.6515,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.4125,
+      "learning_rate": 0.0001,
+      "loss": 7.2849,
+      "loss/crossentropy": 2.0947029411792757,
+      "loss/hidden": 3.359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18155105784535408,
+      "step": 26060
+    },
+    {
+      "epoch": 0.65175,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.3160807291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3504,
+      "loss/crossentropy": 2.0333006374537943,
+      "loss/hidden": 3.40078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21557488106191158,
+      "step": 26070
+    },
+    {
+      "epoch": 0.652,
+      "grad_norm": 66.0,
+      "grad_norm_var": 78.5587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2023,
+      "loss/crossentropy": 2.028635681420565,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18349210238084196,
+      "step": 26080
+    },
+    {
+      "epoch": 0.65225,
+      "grad_norm": 29.875,
+      "grad_norm_var": 81.18723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2882,
+      "loss/crossentropy": 2.110158285498619,
+      "loss/hidden": 3.305078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18722082171589136,
+      "step": 26090
+    },
+    {
+      "epoch": 0.6525,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.225,
+      "learning_rate": 0.0001,
+      "loss": 7.251,
+      "loss/crossentropy": 2.032511693239212,
+      "loss/hidden": 3.329296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20495209330692887,
+      "step": 26100
+    },
+    {
+      "epoch": 0.65275,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.6942057291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3307,
+      "loss/crossentropy": 2.1366796389222147,
+      "loss/hidden": 3.29453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18084107786417009,
+      "step": 26110
+    },
+    {
+      "epoch": 0.653,
+      "grad_norm": 33.0,
+      "grad_norm_var": 1.9559895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2364,
+      "loss/crossentropy": 2.198021276295185,
+      "loss/hidden": 3.334375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1900389553979039,
+      "step": 26120
+    },
+    {
+      "epoch": 0.65325,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.31640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2938,
+      "loss/crossentropy": 2.0711926236748694,
+      "loss/hidden": 3.28203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17610593941062688,
+      "step": 26130
+    },
+    {
+      "epoch": 0.6535,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.4072916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2918,
+      "loss/crossentropy": 2.1544741809368135,
+      "loss/hidden": 3.4359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1978639720007777,
+      "step": 26140
+    },
+    {
+      "epoch": 0.65375,
+      "grad_norm": 28.75,
+      "grad_norm_var": 3.3955729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2798,
+      "loss/crossentropy": 2.0444336414337156,
+      "loss/hidden": 3.42109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19784381221979858,
+      "step": 26150
+    },
+    {
+      "epoch": 0.654,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.7181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.1077,
+      "loss/crossentropy": 1.962803577259183,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17380120740272104,
+      "step": 26160
+    },
+    {
+      "epoch": 0.65425,
+      "grad_norm": 28.75,
+      "grad_norm_var": 10.5541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2623,
+      "loss/crossentropy": 2.0290966272354125,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1842226268723607,
+      "step": 26170
+    },
+    {
+      "epoch": 0.6545,
+      "grad_norm": 31.5,
+      "grad_norm_var": 8.968489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2974,
+      "loss/crossentropy": 2.0671102195978164,
+      "loss/hidden": 3.2921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1779710568487644,
+      "step": 26180
+    },
+    {
+      "epoch": 0.65475,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.0238932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1929,
+      "loss/crossentropy": 1.9803795799613,
+      "loss/hidden": 3.2546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16597625594586135,
+      "step": 26190
+    },
+    {
+      "epoch": 0.655,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.8580729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2966,
+      "loss/crossentropy": 2.0811556324362757,
+      "loss/hidden": 3.396484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1854775669053197,
+      "step": 26200
+    },
+    {
+      "epoch": 0.65525,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.6707682291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2357,
+      "loss/crossentropy": 2.1400867104530334,
+      "loss/hidden": 3.327734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.186828719265759,
+      "step": 26210
+    },
+    {
+      "epoch": 0.6555,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.0072916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3599,
+      "loss/crossentropy": 2.144706717133522,
+      "loss/hidden": 3.33359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19139624712988734,
+      "step": 26220
+    },
+    {
+      "epoch": 0.65575,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.9947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3455,
+      "loss/crossentropy": 1.9842251062393188,
+      "loss/hidden": 3.358203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18394090216606857,
+      "step": 26230
+    },
+    {
+      "epoch": 0.656,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.8645833333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2688,
+      "loss/crossentropy": 2.0962007105350495,
+      "loss/hidden": 3.376953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19022967889904976,
+      "step": 26240
+    },
+    {
+      "epoch": 0.65625,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.3143229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.1723,
+      "loss/crossentropy": 2.0132046937942505,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1787968127988279,
+      "step": 26250
+    },
+    {
+      "epoch": 0.6565,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.4431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3366,
+      "loss/crossentropy": 1.9615045070648194,
+      "loss/hidden": 3.530078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1971464239060879,
+      "step": 26260
+    },
+    {
+      "epoch": 0.65675,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.0830729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.1952,
+      "loss/crossentropy": 2.1030160948634147,
+      "loss/hidden": 3.43828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19630551328882576,
+      "step": 26270
+    },
+    {
+      "epoch": 0.657,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.318489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3209,
+      "loss/crossentropy": 2.2092444196343424,
+      "loss/hidden": 3.28046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18095392240211366,
+      "step": 26280
+    },
+    {
+      "epoch": 0.65725,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.3046223958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2758,
+      "loss/crossentropy": 2.183464777469635,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18903474770486356,
+      "step": 26290
+    },
+    {
+      "epoch": 0.6575,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.9020182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4311,
+      "loss/crossentropy": 2.121826934814453,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18777972608804702,
+      "step": 26300
+    },
+    {
+      "epoch": 0.65775,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.5416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3132,
+      "loss/crossentropy": 1.9226845420897007,
+      "loss/hidden": 3.371484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17548658936284484,
+      "step": 26310
+    },
+    {
+      "epoch": 0.658,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.8861138654567685e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.347,
+      "loss/crossentropy": 2.1671392902731896,
+      "loss/hidden": 3.2453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18836958818137645,
+      "step": 26320
+    },
+    {
+      "epoch": 0.65825,
+      "grad_norm": 30.375,
+      "grad_norm_var": 5.1875,
+      "learning_rate": 0.0001,
+      "loss": 7.2642,
+      "loss/crossentropy": 2.178332243859768,
+      "loss/hidden": 3.2421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17817392963916062,
+      "step": 26330
+    },
+    {
+      "epoch": 0.6585,
+      "grad_norm": 27.875,
+      "grad_norm_var": 1.3184895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2345,
+      "loss/crossentropy": 2.0184032306075097,
+      "loss/hidden": 3.3078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18142256420105696,
+      "step": 26340
+    },
+    {
+      "epoch": 0.65875,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.35390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4358,
+      "loss/crossentropy": 2.070467638969421,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18492308855056763,
+      "step": 26350
+    },
+    {
+      "epoch": 0.659,
+      "grad_norm": 30.5,
+      "grad_norm_var": 37.557291666666664,
+      "learning_rate": 0.0001,
+      "loss": 7.4061,
+      "loss/crossentropy": 2.0928609974682333,
+      "loss/hidden": 3.431640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19751368090510368,
+      "step": 26360
+    },
+    {
+      "epoch": 0.65925,
+      "grad_norm": 31.625,
+      "grad_norm_var": 48.10182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3098,
+      "loss/crossentropy": 2.0748341269791126,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19011560268700123,
+      "step": 26370
+    },
+    {
+      "epoch": 0.6595,
+      "grad_norm": 32.25,
+      "grad_norm_var": 8.820833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3229,
+      "loss/crossentropy": 1.9827906161546707,
+      "loss/hidden": 3.305078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17519382825121282,
+      "step": 26380
+    },
+    {
+      "epoch": 0.65975,
+      "grad_norm": 30.375,
+      "grad_norm_var": 21.3150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3129,
+      "loss/crossentropy": 2.027328287810087,
+      "loss/hidden": 3.359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1863293987698853,
+      "step": 26390
+    },
+    {
+      "epoch": 0.66,
+      "grad_norm": 34.25,
+      "grad_norm_var": 11.154622395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3147,
+      "loss/crossentropy": 2.0382798433303835,
+      "loss/hidden": 3.2234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17127347458153963,
+      "step": 26400
+    },
+    {
+      "epoch": 0.66025,
+      "grad_norm": 29.125,
+      "grad_norm_var": 11.134309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1528,
+      "loss/crossentropy": 2.187176838517189,
+      "loss/hidden": 3.320703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18337102625519036,
+      "step": 26410
+    },
+    {
+      "epoch": 0.6605,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.5205729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.1994,
+      "loss/crossentropy": 1.8949346490204335,
+      "loss/hidden": 3.483984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1746540881693363,
+      "step": 26420
+    },
+    {
+      "epoch": 0.66075,
+      "grad_norm": 29.875,
+      "grad_norm_var": 6.1259765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3588,
+      "loss/crossentropy": 2.1892473474144936,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18937738612294197,
+      "step": 26430
+    },
+    {
+      "epoch": 0.661,
+      "grad_norm": 29.5,
+      "grad_norm_var": 5.56640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3516,
+      "loss/crossentropy": 2.1834142342209817,
+      "loss/hidden": 3.40078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19312054924666883,
+      "step": 26440
+    },
+    {
+      "epoch": 0.66125,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.439583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3007,
+      "loss/crossentropy": 2.1448648869991302,
+      "loss/hidden": 3.483203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2111048873513937,
+      "step": 26450
+    },
+    {
+      "epoch": 0.6615,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.1797421966524677e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3607,
+      "loss/crossentropy": 1.9932497456669807,
+      "loss/hidden": 3.45390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2144283948466182,
+      "step": 26460
+    },
+    {
+      "epoch": 0.66175,
+      "grad_norm": 34.5,
+      "grad_norm_var": 5.677018229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.306,
+      "loss/crossentropy": 2.0623308643698692,
+      "loss/hidden": 3.3484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18320082779973745,
+      "step": 26470
+    },
+    {
+      "epoch": 0.662,
+      "grad_norm": 27.0,
+      "grad_norm_var": 4.14765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3311,
+      "loss/crossentropy": 2.0493450671434403,
+      "loss/hidden": 3.4734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19104604441672562,
+      "step": 26480
+    },
+    {
+      "epoch": 0.66225,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.660872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2653,
+      "loss/crossentropy": 2.031788290292025,
+      "loss/hidden": 3.29609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1821944302879274,
+      "step": 26490
+    },
+    {
+      "epoch": 0.6625,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.8124348958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4374,
+      "loss/crossentropy": 2.1484886214137076,
+      "loss/hidden": 3.4734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20224297530949115,
+      "step": 26500
+    },
+    {
+      "epoch": 0.66275,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.9587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2393,
+      "loss/crossentropy": 2.018560293316841,
+      "loss/hidden": 3.366796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18759384863078593,
+      "step": 26510
+    },
+    {
+      "epoch": 0.663,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.620247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3181,
+      "loss/crossentropy": 2.145674556493759,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18435985408723354,
+      "step": 26520
+    },
+    {
+      "epoch": 0.66325,
+      "grad_norm": 31.125,
+      "grad_norm_var": 7.412239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2507,
+      "loss/crossentropy": 2.0798817604780195,
+      "loss/hidden": 3.341796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18773154616355897,
+      "step": 26530
+    },
+    {
+      "epoch": 0.6635,
+      "grad_norm": 29.5,
+      "grad_norm_var": 7.8681640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2586,
+      "loss/crossentropy": 2.061510816216469,
+      "loss/hidden": 3.271875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19057576302438975,
+      "step": 26540
+    },
+    {
+      "epoch": 0.66375,
+      "grad_norm": 31.125,
+      "grad_norm_var": 6.52890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3062,
+      "loss/crossentropy": 2.072283536195755,
+      "loss/hidden": 3.26328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18644256051629782,
+      "step": 26550
+    },
+    {
+      "epoch": 0.664,
+      "grad_norm": 29.5,
+      "grad_norm_var": 7.605143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2383,
+      "loss/crossentropy": 1.9341054081916809,
+      "loss/hidden": 3.362109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16742698140442372,
+      "step": 26560
+    },
+    {
+      "epoch": 0.66425,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.6518229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.244,
+      "loss/crossentropy": 2.0679861828684807,
+      "loss/hidden": 3.309375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17691615503281355,
+      "step": 26570
+    },
+    {
+      "epoch": 0.6645,
+      "grad_norm": 31.125,
+      "grad_norm_var": 4.612434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3328,
+      "loss/crossentropy": 2.1904489412903785,
+      "loss/hidden": 3.303125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1809101757593453,
+      "step": 26580
+    },
+    {
+      "epoch": 0.66475,
+      "grad_norm": 38.5,
+      "grad_norm_var": 7.606705729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2347,
+      "loss/crossentropy": 1.9506058931350707,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19015372060239316,
+      "step": 26590
+    },
+    {
+      "epoch": 0.665,
+      "grad_norm": 30.0,
+      "grad_norm_var": 6.116080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.1566,
+      "loss/crossentropy": 1.8624847821891308,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1748434605076909,
+      "step": 26600
+    },
+    {
+      "epoch": 0.66525,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.2030598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2837,
+      "loss/crossentropy": 2.1139489620923997,
+      "loss/hidden": 3.26953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1822458105161786,
+      "step": 26610
+    },
+    {
+      "epoch": 0.6655,
+      "grad_norm": 28.875,
+      "grad_norm_var": 46.1978515625,
+      "learning_rate": 0.0001,
+      "loss": 7.1554,
+      "loss/crossentropy": 1.8614851802587509,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16477091368287802,
+      "step": 26620
+    },
+    {
+      "epoch": 0.66575,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.6504557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2609,
+      "loss/crossentropy": 2.2226543337106706,
+      "loss/hidden": 3.301953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1956513339653611,
+      "step": 26630
+    },
+    {
+      "epoch": 0.666,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.601822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2908,
+      "loss/crossentropy": 2.0410634987056255,
+      "loss/hidden": 3.279296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.171825285628438,
+      "step": 26640
+    },
+    {
+      "epoch": 0.66625,
+      "grad_norm": 28.75,
+      "grad_norm_var": 19.392708333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2709,
+      "loss/crossentropy": 2.044853265583515,
+      "loss/hidden": 3.29921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18632043413817884,
+      "step": 26650
+    },
+    {
+      "epoch": 0.6665,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.559375,
+      "learning_rate": 0.0001,
+      "loss": 7.2649,
+      "loss/crossentropy": 2.1507553339004515,
+      "loss/hidden": 3.284375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17937973625957965,
+      "step": 26660
+    },
+    {
+      "epoch": 0.66675,
+      "grad_norm": 33.25,
+      "grad_norm_var": 3.41875,
+      "learning_rate": 0.0001,
+      "loss": 7.3469,
+      "loss/crossentropy": 2.1353399351239206,
+      "loss/hidden": 3.253125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17566792499274014,
+      "step": 26670
+    },
+    {
+      "epoch": 0.667,
+      "grad_norm": 28.125,
+      "grad_norm_var": 2.9119140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2886,
+      "loss/crossentropy": 2.0442838847637175,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1838838504627347,
+      "step": 26680
+    },
+    {
+      "epoch": 0.66725,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.5166015625,
+      "learning_rate": 0.0001,
+      "loss": 7.268,
+      "loss/crossentropy": 2.0146602764725685,
+      "loss/hidden": 3.432421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.189494782499969,
+      "step": 26690
+    },
+    {
+      "epoch": 0.6675,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.2447916666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2655,
+      "loss/crossentropy": 2.062611496448517,
+      "loss/hidden": 3.42578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20330444928258656,
+      "step": 26700
+    },
+    {
+      "epoch": 0.66775,
+      "grad_norm": 28.375,
+      "grad_norm_var": 3.17265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3095,
+      "loss/crossentropy": 2.1450613498687745,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18933778218925,
+      "step": 26710
+    },
+    {
+      "epoch": 0.668,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.965625,
+      "learning_rate": 0.0001,
+      "loss": 7.2676,
+      "loss/crossentropy": 1.9853663839399815,
+      "loss/hidden": 3.250390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17568059861660004,
+      "step": 26720
+    },
+    {
+      "epoch": 0.66825,
+      "grad_norm": 29.75,
+      "grad_norm_var": 115517274069465.19,
+      "learning_rate": 0.0001,
+      "loss": 7.3871,
+      "loss/crossentropy": 2.2481682300567627,
+      "loss/hidden": 3.358984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19613477736711502,
+      "step": 26730
+    },
+    {
+      "epoch": 0.6685,
+      "grad_norm": 29.875,
+      "grad_norm_var": 115517270039004.08,
+      "learning_rate": 0.0001,
+      "loss": 7.3836,
+      "loss/crossentropy": 2.221486309170723,
+      "loss/hidden": 3.33359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19494929853826762,
+      "step": 26740
+    },
+    {
+      "epoch": 0.66875,
+      "grad_norm": 27.875,
+      "grad_norm_var": 1.67890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2099,
+      "loss/crossentropy": 2.0120978243649006,
+      "loss/hidden": 3.4109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16912673255428673,
+      "step": 26750
+    },
+    {
+      "epoch": 0.669,
+      "grad_norm": 28.0,
+      "grad_norm_var": 3.424934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2854,
+      "loss/crossentropy": 2.1522657334804536,
+      "loss/hidden": 3.284375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17450021374970676,
+      "step": 26760
+    },
+    {
+      "epoch": 0.66925,
+      "grad_norm": 32.5,
+      "grad_norm_var": 10.490625,
+      "learning_rate": 0.0001,
+      "loss": 7.4043,
+      "loss/crossentropy": 2.03169424161315,
+      "loss/hidden": 3.3890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18973638992756606,
+      "step": 26770
+    },
+    {
+      "epoch": 0.6695,
+      "grad_norm": 40.5,
+      "grad_norm_var": 13.8400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3005,
+      "loss/crossentropy": 2.010363797843456,
+      "loss/hidden": 3.431640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2009760271757841,
+      "step": 26780
+    },
+    {
+      "epoch": 0.66975,
+      "grad_norm": 29.875,
+      "grad_norm_var": 9.314518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3094,
+      "loss/crossentropy": 2.0870240703225136,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19831706527620555,
+      "step": 26790
+    },
+    {
+      "epoch": 0.67,
+      "grad_norm": 34.0,
+      "grad_norm_var": 2.0208333333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3591,
+      "loss/crossentropy": 1.9903123408555985,
+      "loss/hidden": 3.4703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18768544010818006,
+      "step": 26800
+    },
+    {
+      "epoch": 0.67025,
+      "grad_norm": 30.75,
+      "grad_norm_var": 5.1962890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3379,
+      "loss/crossentropy": 1.9887514635920525,
+      "loss/hidden": 3.42734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18731831386685371,
+      "step": 26810
+    },
+    {
+      "epoch": 0.6705,
+      "grad_norm": 30.875,
+      "grad_norm_var": 5.4275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2099,
+      "loss/crossentropy": 2.2334194481372833,
+      "loss/hidden": 3.294921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18677575308829547,
+      "step": 26820
+    },
+    {
+      "epoch": 0.67075,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.9624348958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2166,
+      "loss/crossentropy": 2.09789634346962,
+      "loss/hidden": 3.352734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1792703490704298,
+      "step": 26830
+    },
+    {
+      "epoch": 0.671,
+      "grad_norm": 28.75,
+      "grad_norm_var": 1.0145833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2917,
+      "loss/crossentropy": 2.083233179152012,
+      "loss/hidden": 3.42734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1986743938177824,
+      "step": 26840
+    },
+    {
+      "epoch": 0.67125,
+      "grad_norm": 37.25,
+      "grad_norm_var": 23.931184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2443,
+      "loss/crossentropy": 2.0056370086967945,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1743487582542002,
+      "step": 26850
+    },
+    {
+      "epoch": 0.6715,
+      "grad_norm": 30.0,
+      "grad_norm_var": 21.345768229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2698,
+      "loss/crossentropy": 2.222350174188614,
+      "loss/hidden": 3.3890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.199802653118968,
+      "step": 26860
+    },
+    {
+      "epoch": 0.67175,
+      "grad_norm": 30.5,
+      "grad_norm_var": 106.80930989583334,
+      "learning_rate": 0.0001,
+      "loss": 7.3626,
+      "loss/crossentropy": 2.0834795638918875,
+      "loss/hidden": 3.463671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18940293360501528,
+      "step": 26870
+    },
+    {
+      "epoch": 0.672,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.638541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3212,
+      "loss/crossentropy": 2.131199154257774,
+      "loss/hidden": 3.406640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19130074717104434,
+      "step": 26880
+    },
+    {
+      "epoch": 0.67225,
+      "grad_norm": 7314866176.0,
+      "grad_norm_var": 3.3442041705028127e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.269,
+      "loss/crossentropy": 2.047175918519497,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.174586620926857,
+      "step": 26890
+    },
+    {
+      "epoch": 0.6725,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.344204168613139e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3649,
+      "loss/crossentropy": 2.140024873614311,
+      "loss/hidden": 3.348828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1875780440866947,
+      "step": 26900
+    },
+    {
+      "epoch": 0.67275,
+      "grad_norm": 30.375,
+      "grad_norm_var": 19.369791666666668,
+      "learning_rate": 0.0001,
+      "loss": 7.2309,
+      "loss/crossentropy": 2.0859541945159434,
+      "loss/hidden": 3.3109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17924507465213538,
+      "step": 26910
+    },
+    {
+      "epoch": 0.673,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.1056640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3053,
+      "loss/crossentropy": 1.9552275583148002,
+      "loss/hidden": 3.28203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16615418754518033,
+      "step": 26920
+    },
+    {
+      "epoch": 0.67325,
+      "grad_norm": 29.125,
+      "grad_norm_var": 9.1228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3161,
+      "loss/crossentropy": 2.051589508354664,
+      "loss/hidden": 3.43359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1953272173181176,
+      "step": 26930
+    },
+    {
+      "epoch": 0.6735,
+      "grad_norm": 33.25,
+      "grad_norm_var": 12.4322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2163,
+      "loss/crossentropy": 2.039374630153179,
+      "loss/hidden": 3.320703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1790797933936119,
+      "step": 26940
+    },
+    {
+      "epoch": 0.67375,
+      "grad_norm": 29.5,
+      "grad_norm_var": 10.328580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2611,
+      "loss/crossentropy": 2.135941173136234,
+      "loss/hidden": 3.310546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18610633611679078,
+      "step": 26950
+    },
+    {
+      "epoch": 0.674,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.2393229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2542,
+      "loss/crossentropy": 2.061223492026329,
+      "loss/hidden": 3.237109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1693508742377162,
+      "step": 26960
+    },
+    {
+      "epoch": 0.67425,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.8309895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3494,
+      "loss/crossentropy": 2.034488780796528,
+      "loss/hidden": 3.362109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18367824982851744,
+      "step": 26970
+    },
+    {
+      "epoch": 0.6745,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.6082682291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2714,
+      "loss/crossentropy": 2.127776539325714,
+      "loss/hidden": 3.454296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19541560169309377,
+      "step": 26980
+    },
+    {
+      "epoch": 0.67475,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.8431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2163,
+      "loss/crossentropy": 1.9815926529467105,
+      "loss/hidden": 3.291796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.187463073246181,
+      "step": 26990
+    },
+    {
+      "epoch": 0.675,
+      "grad_norm": 29.875,
+      "grad_norm_var": 12.0572265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3076,
+      "loss/crossentropy": 2.025939218699932,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18148638047277926,
+      "step": 27000
+    },
+    {
+      "epoch": 0.67525,
+      "grad_norm": 28.25,
+      "grad_norm_var": 2.4364583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2504,
+      "loss/crossentropy": 2.1637841314077377,
+      "loss/hidden": 3.25625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1768895300105214,
+      "step": 27010
+    },
+    {
+      "epoch": 0.6755,
+      "grad_norm": 27.375,
+      "grad_norm_var": 3.078125,
+      "learning_rate": 0.0001,
+      "loss": 7.2796,
+      "loss/crossentropy": 2.167166344821453,
+      "loss/hidden": 3.372265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20178660629317163,
+      "step": 27020
+    },
+    {
+      "epoch": 0.67575,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.6186848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3119,
+      "loss/crossentropy": 2.0646814957261084,
+      "loss/hidden": 3.433984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19334694920107723,
+      "step": 27030
+    },
+    {
+      "epoch": 0.676,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.6322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2636,
+      "loss/crossentropy": 2.04083681628108,
+      "loss/hidden": 3.30390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17480793623253704,
+      "step": 27040
+    },
+    {
+      "epoch": 0.67625,
+      "grad_norm": 28.125,
+      "grad_norm_var": 2.957246447870105e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3215,
+      "loss/crossentropy": 2.062483212351799,
+      "loss/hidden": 3.33359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1790636266581714,
+      "step": 27050
+    },
+    {
+      "epoch": 0.6765,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.9572464473757015e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2078,
+      "loss/crossentropy": 2.1033657416701317,
+      "loss/hidden": 3.312890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1807030899450183,
+      "step": 27060
+    },
+    {
+      "epoch": 0.67675,
+      "grad_norm": 28.875,
+      "grad_norm_var": 5.155208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2486,
+      "loss/crossentropy": 1.976552402228117,
+      "loss/hidden": 3.340625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17463484006002544,
+      "step": 27070
+    },
+    {
+      "epoch": 0.677,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.590625,
+      "learning_rate": 0.0001,
+      "loss": 7.2763,
+      "loss/crossentropy": 2.1383912026882173,
+      "loss/hidden": 3.283984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1793740052729845,
+      "step": 27080
+    },
+    {
+      "epoch": 0.67725,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.7791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3335,
+      "loss/crossentropy": 2.183702952414751,
+      "loss/hidden": 3.391015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20361373238265515,
+      "step": 27090
+    },
+    {
+      "epoch": 0.6775,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.4552083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1603,
+      "loss/crossentropy": 2.2112974256277083,
+      "loss/hidden": 3.250390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17802942357957363,
+      "step": 27100
+    },
+    {
+      "epoch": 0.67775,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.090625,
+      "learning_rate": 0.0001,
+      "loss": 7.282,
+      "loss/crossentropy": 2.1228010304272176,
+      "loss/hidden": 3.36171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1823104264214635,
+      "step": 27110
+    },
+    {
+      "epoch": 0.678,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.84765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2531,
+      "loss/crossentropy": 2.0709615498781204,
+      "loss/hidden": 3.299609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.171175280213356,
+      "step": 27120
+    },
+    {
+      "epoch": 0.67825,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.8655598958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3062,
+      "loss/crossentropy": 2.2544207006692885,
+      "loss/hidden": 3.384765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20476784072816373,
+      "step": 27130
+    },
+    {
+      "epoch": 0.6785,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.9958333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.269,
+      "loss/crossentropy": 2.044566038250923,
+      "loss/hidden": 3.416015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18673994652926923,
+      "step": 27140
+    },
+    {
+      "epoch": 0.67875,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.787239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3861,
+      "loss/crossentropy": 2.076129969954491,
+      "loss/hidden": 3.269921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17596831321716308,
+      "step": 27150
+    },
+    {
+      "epoch": 0.679,
+      "grad_norm": 30.375,
+      "grad_norm_var": 4.278059895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4392,
+      "loss/crossentropy": 2.0484411746263502,
+      "loss/hidden": 3.398828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1874468218535185,
+      "step": 27160
+    },
+    {
+      "epoch": 0.67925,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.25390625,
+      "learning_rate": 0.0001,
+      "loss": 7.1788,
+      "loss/crossentropy": 2.180029886960983,
+      "loss/hidden": 3.38984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20117557551711798,
+      "step": 27170
+    },
+    {
+      "epoch": 0.6795,
+      "grad_norm": 31.625,
+      "grad_norm_var": 5.522330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2891,
+      "loss/crossentropy": 1.8975796066224575,
+      "loss/hidden": 3.30703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17547342590987683,
+      "step": 27180
+    },
+    {
+      "epoch": 0.67975,
+      "grad_norm": 30.375,
+      "grad_norm_var": 11.6009765625,
+      "learning_rate": 0.0001,
+      "loss": 7.1943,
+      "loss/crossentropy": 2.1247182063758374,
+      "loss/hidden": 3.315625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17981126178056,
+      "step": 27190
+    },
+    {
+      "epoch": 0.68,
+      "grad_norm": 28.25,
+      "grad_norm_var": 8.919205729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2694,
+      "loss/crossentropy": 2.037150639295578,
+      "loss/hidden": 3.262109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17851810567080975,
+      "step": 27200
+    },
+    {
+      "epoch": 0.68025,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.218489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2526,
+      "loss/crossentropy": 1.9110659010708333,
+      "loss/hidden": 3.5203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19786193277686834,
+      "step": 27210
+    },
+    {
+      "epoch": 0.6805,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.8385416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2515,
+      "loss/crossentropy": 2.2368867814540865,
+      "loss/hidden": 3.2171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1798218859359622,
+      "step": 27220
+    },
+    {
+      "epoch": 0.68075,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.54140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2991,
+      "loss/crossentropy": 2.1260806649923323,
+      "loss/hidden": 3.43359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20750290621072054,
+      "step": 27230
+    },
+    {
+      "epoch": 0.681,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.755989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3501,
+      "loss/crossentropy": 2.031808337569237,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18845813032239675,
+      "step": 27240
+    },
+    {
+      "epoch": 0.68125,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.9942057291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2698,
+      "loss/crossentropy": 2.1050592973828315,
+      "loss/hidden": 3.322265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18391012214124203,
+      "step": 27250
+    },
+    {
+      "epoch": 0.6815,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.36640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3684,
+      "loss/crossentropy": 2.1474289670586586,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1923911923542619,
+      "step": 27260
+    },
+    {
+      "epoch": 0.68175,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.9858723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2966,
+      "loss/crossentropy": 2.0553861789405348,
+      "loss/hidden": 3.321484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17631796486675738,
+      "step": 27270
+    },
+    {
+      "epoch": 0.682,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.138541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.27,
+      "loss/crossentropy": 1.9470392048358918,
+      "loss/hidden": 3.433203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1823154903948307,
+      "step": 27280
+    },
+    {
+      "epoch": 0.68225,
+      "grad_norm": 29.375,
+      "grad_norm_var": 6.4525390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3825,
+      "loss/crossentropy": 2.1857406944036484,
+      "loss/hidden": 3.442578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19291127175092698,
+      "step": 27290
+    },
+    {
+      "epoch": 0.6825,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.48515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3561,
+      "loss/crossentropy": 2.227454760670662,
+      "loss/hidden": 3.362109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18692610040307045,
+      "step": 27300
+    },
+    {
+      "epoch": 0.68275,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.13515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2776,
+      "loss/crossentropy": 1.9948305085301399,
+      "loss/hidden": 3.390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1988155659288168,
+      "step": 27310
+    },
+    {
+      "epoch": 0.683,
+      "grad_norm": 28.875,
+      "grad_norm_var": 1.8541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2481,
+      "loss/crossentropy": 2.1666594982147216,
+      "loss/hidden": 3.326953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18878179658204317,
+      "step": 27320
+    },
+    {
+      "epoch": 0.68325,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.687239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3854,
+      "loss/crossentropy": 2.1071100682020187,
+      "loss/hidden": 3.30625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17986488938331605,
+      "step": 27330
+    },
+    {
+      "epoch": 0.6835,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.3889973958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.235,
+      "loss/crossentropy": 2.029652649909258,
+      "loss/hidden": 3.409765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1780470222234726,
+      "step": 27340
+    },
+    {
+      "epoch": 0.68375,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.155143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3103,
+      "loss/crossentropy": 1.9602101355791093,
+      "loss/hidden": 3.35859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18400138095021248,
+      "step": 27350
+    },
+    {
+      "epoch": 0.684,
+      "grad_norm": 29.375,
+      "grad_norm_var": 0.45598958333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.213,
+      "loss/crossentropy": 2.106926290690899,
+      "loss/hidden": 3.305078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17419682666659356,
+      "step": 27360
+    },
+    {
+      "epoch": 0.68425,
+      "grad_norm": 29.75,
+      "grad_norm_var": 0.44973958333333336,
+      "learning_rate": 0.0001,
+      "loss": 7.2879,
+      "loss/crossentropy": 2.237940040230751,
+      "loss/hidden": 3.3515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1868164110928774,
+      "step": 27370
+    },
+    {
+      "epoch": 0.6845,
+      "grad_norm": 36.25,
+      "grad_norm_var": 2.9947916666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4911,
+      "loss/crossentropy": 2.2364405103027822,
+      "loss/hidden": 3.362109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1962603386491537,
+      "step": 27380
+    },
+    {
+      "epoch": 0.68475,
+      "grad_norm": 32.75,
+      "grad_norm_var": 7.664583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3546,
+      "loss/crossentropy": 2.125661809742451,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17499309908598662,
+      "step": 27390
+    },
+    {
+      "epoch": 0.685,
+      "grad_norm": 29.125,
+      "grad_norm_var": 7.216666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2694,
+      "loss/crossentropy": 2.17802165299654,
+      "loss/hidden": 3.26953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18032563328742982,
+      "step": 27400
+    },
+    {
+      "epoch": 0.68525,
+      "grad_norm": 28.625,
+      "grad_norm_var": 2.5254557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2098,
+      "loss/crossentropy": 2.0895639203488825,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17885948363691567,
+      "step": 27410
+    },
+    {
+      "epoch": 0.6855,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.098893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2392,
+      "loss/crossentropy": 2.0427750945091248,
+      "loss/hidden": 3.389453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.184170157648623,
+      "step": 27420
+    },
+    {
+      "epoch": 0.68575,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.5853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4218,
+      "loss/crossentropy": 2.2016945309937,
+      "loss/hidden": 3.275390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1827758725732565,
+      "step": 27430
+    },
+    {
+      "epoch": 0.686,
+      "grad_norm": 33.25,
+      "grad_norm_var": 1.7332682291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2304,
+      "loss/crossentropy": 1.969399856030941,
+      "loss/hidden": 3.338671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17159626744687556,
+      "step": 27440
+    },
+    {
+      "epoch": 0.68625,
+      "grad_norm": 34.5,
+      "grad_norm_var": 22.256184895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.2665,
+      "loss/crossentropy": 2.0050867199897766,
+      "loss/hidden": 3.39921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18022035136818887,
+      "step": 27450
+    },
+    {
+      "epoch": 0.6865,
+      "grad_norm": 27.625,
+      "grad_norm_var": 22.336458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2446,
+      "loss/crossentropy": 1.9901010252535343,
+      "loss/hidden": 3.275390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17804777259007096,
+      "step": 27460
+    },
+    {
+      "epoch": 0.68675,
+      "grad_norm": 28.125,
+      "grad_norm_var": 12.424739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2441,
+      "loss/crossentropy": 2.031005633622408,
+      "loss/hidden": 3.290234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.168888134136796,
+      "step": 27470
+    },
+    {
+      "epoch": 0.687,
+      "grad_norm": 29.0,
+      "grad_norm_var": 0.8817057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2623,
+      "loss/crossentropy": 2.1447826243937014,
+      "loss/hidden": 3.35859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18581478726118802,
+      "step": 27480
+    },
+    {
+      "epoch": 0.68725,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.9895833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3027,
+      "loss/crossentropy": 2.2375788524746896,
+      "loss/hidden": 3.38828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19884238932281734,
+      "step": 27490
+    },
+    {
+      "epoch": 0.6875,
+      "grad_norm": 30.875,
+      "grad_norm_var": 10.562239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3156,
+      "loss/crossentropy": 1.9248571954667568,
+      "loss/hidden": 3.317578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16993207540363073,
+      "step": 27500
+    },
+    {
+      "epoch": 0.68775,
+      "grad_norm": 29.125,
+      "grad_norm_var": 6.201041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1426,
+      "loss/crossentropy": 2.1742107778787614,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19332415759563445,
+      "step": 27510
+    },
+    {
+      "epoch": 0.688,
+      "grad_norm": 28.625,
+      "grad_norm_var": 2.443489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1899,
+      "loss/crossentropy": 2.064833169430494,
+      "loss/hidden": 3.226171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1686219139955938,
+      "step": 27520
+    },
+    {
+      "epoch": 0.68825,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.308333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2861,
+      "loss/crossentropy": 2.2087809190154077,
+      "loss/hidden": 3.320703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18179803155362606,
+      "step": 27530
+    },
+    {
+      "epoch": 0.6885,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.3559895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2698,
+      "loss/crossentropy": 2.1848205491900443,
+      "loss/hidden": 3.308203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18062485102564096,
+      "step": 27540
+    },
+    {
+      "epoch": 0.68875,
+      "grad_norm": 28.375,
+      "grad_norm_var": 2.5254557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.1884,
+      "loss/crossentropy": 2.10894877538085,
+      "loss/hidden": 3.278515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18315895553678274,
+      "step": 27550
+    },
+    {
+      "epoch": 0.689,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.7504557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2776,
+      "loss/crossentropy": 1.9970835141837597,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1800586089491844,
+      "step": 27560
+    },
+    {
+      "epoch": 0.68925,
+      "grad_norm": 63.0,
+      "grad_norm_var": 446.81223958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2313,
+      "loss/crossentropy": 2.108223394304514,
+      "loss/hidden": 3.310546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18529674978926777,
+      "step": 27570
+    },
+    {
+      "epoch": 0.6895,
+      "grad_norm": 27.75,
+      "grad_norm_var": 229.2228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.311,
+      "loss/crossentropy": 2.140768423676491,
+      "loss/hidden": 3.317578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18569578658789396,
+      "step": 27580
+    },
+    {
+      "epoch": 0.68975,
+      "grad_norm": 30.0,
+      "grad_norm_var": 141.71608072916666,
+      "learning_rate": 0.0001,
+      "loss": 7.1009,
+      "loss/crossentropy": 2.080308865010738,
+      "loss/hidden": 3.379296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17526689525693656,
+      "step": 27590
+    },
+    {
+      "epoch": 0.69,
+      "grad_norm": 47.5,
+      "grad_norm_var": 127.42649739583334,
+      "learning_rate": 0.0001,
+      "loss": 7.1538,
+      "loss/crossentropy": 1.9624833583831787,
+      "loss/hidden": 3.242578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17015852518379687,
+      "step": 27600
+    },
+    {
+      "epoch": 0.69025,
+      "grad_norm": 28.0,
+      "grad_norm_var": 107.753125,
+      "learning_rate": 0.0001,
+      "loss": 7.2956,
+      "loss/crossentropy": 2.0645004503428934,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19667823929339648,
+      "step": 27610
+    },
+    {
+      "epoch": 0.6905,
+      "grad_norm": 30.5,
+      "grad_norm_var": 73.18723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2826,
+      "loss/crossentropy": 2.18245909512043,
+      "loss/hidden": 3.3125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1764837397262454,
+      "step": 27620
+    },
+    {
+      "epoch": 0.69075,
+      "grad_norm": 54.5,
+      "grad_norm_var": 85.95774739583334,
+      "learning_rate": 0.0001,
+      "loss": 7.1622,
+      "loss/crossentropy": 2.022562848776579,
+      "loss/hidden": 3.396484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17506722770631314,
+      "step": 27630
+    },
+    {
+      "epoch": 0.691,
+      "grad_norm": 29.25,
+      "grad_norm_var": 62.54055989583333,
+      "learning_rate": 0.0001,
+      "loss": 7.3102,
+      "loss/crossentropy": 2.039566272497177,
+      "loss/hidden": 3.392578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18705343939363955,
+      "step": 27640
+    },
+    {
+      "epoch": 0.69125,
+      "grad_norm": 31.625,
+      "grad_norm_var": 86.08639322916666,
+      "learning_rate": 0.0001,
+      "loss": 7.1129,
+      "loss/crossentropy": 2.006290066242218,
+      "loss/hidden": 3.28125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17148794345557689,
+      "step": 27650
+    },
+    {
+      "epoch": 0.6915,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.344204167188264e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.1452,
+      "loss/crossentropy": 2.08501241505146,
+      "loss/hidden": 3.43046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1790543705224991,
+      "step": 27660
+    },
+    {
+      "epoch": 0.69175,
+      "grad_norm": 30.25,
+      "grad_norm_var": 75.49479166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.226,
+      "loss/crossentropy": 2.061854050308466,
+      "loss/hidden": 3.241015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18336585741490125,
+      "step": 27670
+    },
+    {
+      "epoch": 0.692,
+      "grad_norm": 44.75,
+      "grad_norm_var": 54.40390625,
+      "learning_rate": 0.0001,
+      "loss": 7.192,
+      "loss/crossentropy": 2.12304325401783,
+      "loss/hidden": 3.332421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19009512551128865,
+      "step": 27680
+    },
+    {
+      "epoch": 0.69225,
+      "grad_norm": 29.625,
+      "grad_norm_var": 43.609309895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2723,
+      "loss/crossentropy": 2.082204009592533,
+      "loss/hidden": 3.34609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18473013881593942,
+      "step": 27690
+    },
+    {
+      "epoch": 0.6925,
+      "grad_norm": 27.125,
+      "grad_norm_var": 34.3587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2277,
+      "loss/crossentropy": 2.07097297757864,
+      "loss/hidden": 3.36484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1826794480904937,
+      "step": 27700
+    },
+    {
+      "epoch": 0.69275,
+      "grad_norm": 40.0,
+      "grad_norm_var": 27.279622395833332,
+      "learning_rate": 0.0001,
+      "loss": 7.2755,
+      "loss/crossentropy": 2.029601737856865,
+      "loss/hidden": 3.4109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1799439148977399,
+      "step": 27710
+    },
+    {
+      "epoch": 0.693,
+      "grad_norm": 39.25,
+      "grad_norm_var": 17.6291015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2338,
+      "loss/crossentropy": 2.0642464771866798,
+      "loss/hidden": 3.241015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1682391857728362,
+      "step": 27720
+    },
+    {
+      "epoch": 0.69325,
+      "grad_norm": 32.25,
+      "grad_norm_var": 17.788997395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.1999,
+      "loss/crossentropy": 1.9560820698738097,
+      "loss/hidden": 3.362109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18492446821182967,
+      "step": 27730
+    },
+    {
+      "epoch": 0.6935,
+      "grad_norm": 29.375,
+      "grad_norm_var": 19.187239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2068,
+      "loss/crossentropy": 2.151790848374367,
+      "loss/hidden": 3.37109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18812405075877905,
+      "step": 27740
+    },
+    {
+      "epoch": 0.69375,
+      "grad_norm": 33.25,
+      "grad_norm_var": 25.820572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2683,
+      "loss/crossentropy": 2.027320671826601,
+      "loss/hidden": 3.394140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1938678870908916,
+      "step": 27750
+    },
+    {
+      "epoch": 0.694,
+      "grad_norm": 35.75,
+      "grad_norm_var": 13.291080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2137,
+      "loss/crossentropy": 2.222668908536434,
+      "loss/hidden": 3.212890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1686727624386549,
+      "step": 27760
+    },
+    {
+      "epoch": 0.69425,
+      "grad_norm": 33.5,
+      "grad_norm_var": 17.365559895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.2981,
+      "loss/crossentropy": 2.0196828715503217,
+      "loss/hidden": 3.327734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17830958236008881,
+      "step": 27770
+    },
+    {
+      "epoch": 0.6945,
+      "grad_norm": 31.25,
+      "grad_norm_var": 13.575,
+      "learning_rate": 0.0001,
+      "loss": 7.2654,
+      "loss/crossentropy": 2.0276674374938013,
+      "loss/hidden": 3.34296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1810338616371155,
+      "step": 27780
+    },
+    {
+      "epoch": 0.69475,
+      "grad_norm": 38.5,
+      "grad_norm_var": 1.0995116099021482e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.1378,
+      "loss/crossentropy": 1.9873436450958253,
+      "loss/hidden": 3.31953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1692363377660513,
+      "step": 27790
+    },
+    {
+      "epoch": 0.695,
+      "grad_norm": 30.25,
+      "grad_norm_var": 19.480989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1259,
+      "loss/crossentropy": 2.0556870639324187,
+      "loss/hidden": 3.336328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1810257313773036,
+      "step": 27800
+    },
+    {
+      "epoch": 0.69525,
+      "grad_norm": 31.375,
+      "grad_norm_var": 18.948372395833335,
+      "learning_rate": 0.0001,
+      "loss": 7.2655,
+      "loss/crossentropy": 2.047024370729923,
+      "loss/hidden": 3.299609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.176701095700264,
+      "step": 27810
+    },
+    {
+      "epoch": 0.6955,
+      "grad_norm": 28.375,
+      "grad_norm_var": 12.296875,
+      "learning_rate": 0.0001,
+      "loss": 7.2012,
+      "loss/crossentropy": 1.9956993229687214,
+      "loss/hidden": 3.25078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1679278828203678,
+      "step": 27820
+    },
+    {
+      "epoch": 0.69575,
+      "grad_norm": 29.875,
+      "grad_norm_var": 10.91015625,
+      "learning_rate": 0.0001,
+      "loss": 7.1434,
+      "loss/crossentropy": 1.9673665076494218,
+      "loss/hidden": 3.403125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1817054906859994,
+      "step": 27830
+    },
+    {
+      "epoch": 0.696,
+      "grad_norm": 27.25,
+      "grad_norm_var": 65.115625,
+      "learning_rate": 0.0001,
+      "loss": 7.1678,
+      "loss/crossentropy": 2.110173484683037,
+      "loss/hidden": 3.299609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17987647745758295,
+      "step": 27840
+    },
+    {
+      "epoch": 0.69625,
+      "grad_norm": 27.0,
+      "grad_norm_var": 80.79375,
+      "learning_rate": 0.0001,
+      "loss": 7.1452,
+      "loss/crossentropy": 2.066749632358551,
+      "loss/hidden": 3.41484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18724118024110795,
+      "step": 27850
+    },
+    {
+      "epoch": 0.6965,
+      "grad_norm": 27.625,
+      "grad_norm_var": 37.25182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1831,
+      "loss/crossentropy": 1.8799223847687245,
+      "loss/hidden": 3.3375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1720001684501767,
+      "step": 27860
+    },
+    {
+      "epoch": 0.69675,
+      "grad_norm": 31.875,
+      "grad_norm_var": 6.83515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2257,
+      "loss/crossentropy": 1.984929259866476,
+      "loss/hidden": 3.315234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18079593311995268,
+      "step": 27870
+    },
+    {
+      "epoch": 0.697,
+      "grad_norm": 28.375,
+      "grad_norm_var": 12.743489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1256,
+      "loss/crossentropy": 2.1968638002872467,
+      "loss/hidden": 3.26328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17731832768768072,
+      "step": 27880
+    },
+    {
+      "epoch": 0.69725,
+      "grad_norm": 29.0,
+      "grad_norm_var": 9.189583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.181,
+      "loss/crossentropy": 1.9869700893759727,
+      "loss/hidden": 3.27734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18951656930148603,
+      "step": 27890
+    },
+    {
+      "epoch": 0.6975,
+      "grad_norm": 29.5,
+      "grad_norm_var": 12.85390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2834,
+      "loss/crossentropy": 2.1294341668486596,
+      "loss/hidden": 3.38359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18512535598129035,
+      "step": 27900
+    },
+    {
+      "epoch": 0.69775,
+      "grad_norm": 29.625,
+      "grad_norm_var": 6.847916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.175,
+      "loss/crossentropy": 2.1979243993759154,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2182347685098648,
+      "step": 27910
+    },
+    {
+      "epoch": 0.698,
+      "grad_norm": 29.0,
+      "grad_norm_var": 8.465625,
+      "learning_rate": 0.0001,
+      "loss": 7.2304,
+      "loss/crossentropy": 1.9434294357895852,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1738783059641719,
+      "step": 27920
+    },
+    {
+      "epoch": 0.69825,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.443489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1309,
+      "loss/crossentropy": 1.9772110790014268,
+      "loss/hidden": 3.26328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16446346305310727,
+      "step": 27930
+    },
+    {
+      "epoch": 0.6985,
+      "grad_norm": 31.625,
+      "grad_norm_var": 7.475,
+      "learning_rate": 0.0001,
+      "loss": 7.2323,
+      "loss/crossentropy": 2.046364203095436,
+      "loss/hidden": 3.24453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16903265602886677,
+      "step": 27940
+    },
+    {
+      "epoch": 0.69875,
+      "grad_norm": 29.5,
+      "grad_norm_var": 4.887434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3053,
+      "loss/crossentropy": 2.240569099038839,
+      "loss/hidden": 3.337890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18274808302521706,
+      "step": 27950
+    },
+    {
+      "epoch": 0.699,
+      "grad_norm": 34.0,
+      "grad_norm_var": 7.443684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3101,
+      "loss/crossentropy": 2.044392989575863,
+      "loss/hidden": 3.4140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18658852139487864,
+      "step": 27960
+    },
+    {
+      "epoch": 0.69925,
+      "grad_norm": 31.0,
+      "grad_norm_var": 9.6166015625,
+      "learning_rate": 0.0001,
+      "loss": 7.359,
+      "loss/crossentropy": 2.178907059133053,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1918690962716937,
+      "step": 27970
+    },
+    {
+      "epoch": 0.6995,
+      "grad_norm": 30.875,
+      "grad_norm_var": 6.7806640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3059,
+      "loss/crossentropy": 2.07792577072978,
+      "loss/hidden": 3.40703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1873043139465153,
+      "step": 27980
+    },
+    {
+      "epoch": 0.69975,
+      "grad_norm": 29.75,
+      "grad_norm_var": 5.643489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3217,
+      "loss/crossentropy": 1.9168295338749886,
+      "loss/hidden": 3.4546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18331131394952535,
+      "step": 27990
+    },
+    {
+      "epoch": 0.7,
+      "grad_norm": 30.625,
+      "grad_norm_var": 4.960416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3017,
+      "loss/crossentropy": 1.9929687798023223,
+      "loss/hidden": 3.482421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21256182622164488,
+      "step": 28000
+    },
+    {
+      "epoch": 0.70025,
+      "grad_norm": 28.25,
+      "grad_norm_var": 227.75625,
+      "learning_rate": 0.0001,
+      "loss": 7.2574,
+      "loss/crossentropy": 2.0593857660889627,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17820099778473378,
+      "step": 28010
+    },
+    {
+      "epoch": 0.7005,
+      "grad_norm": 31.25,
+      "grad_norm_var": 224.015625,
+      "learning_rate": 0.0001,
+      "loss": 7.412,
+      "loss/crossentropy": 2.174033749103546,
+      "loss/hidden": 3.301171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1898373905569315,
+      "step": 28020
+    },
+    {
+      "epoch": 0.70075,
+      "grad_norm": 28.375,
+      "grad_norm_var": 4.4181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.1913,
+      "loss/crossentropy": 2.0483900040388106,
+      "loss/hidden": 3.38125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17332715671509505,
+      "step": 28030
+    },
+    {
+      "epoch": 0.701,
+      "grad_norm": 29.5,
+      "grad_norm_var": 5.629166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2349,
+      "loss/crossentropy": 2.177481460571289,
+      "loss/hidden": 3.284375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1783738609403372,
+      "step": 28040
+    },
+    {
+      "epoch": 0.70125,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.4809895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3133,
+      "loss/crossentropy": 2.099740183353424,
+      "loss/hidden": 3.358984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2023984882980585,
+      "step": 28050
+    },
+    {
+      "epoch": 0.7015,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.1051432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2455,
+      "loss/crossentropy": 1.9891198098659515,
+      "loss/hidden": 3.476953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1935951853170991,
+      "step": 28060
+    },
+    {
+      "epoch": 0.70175,
+      "grad_norm": 30.5,
+      "grad_norm_var": 23.0619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2426,
+      "loss/crossentropy": 1.9753205917775631,
+      "loss/hidden": 3.33046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17461898345500232,
+      "step": 28070
+    },
+    {
+      "epoch": 0.702,
+      "grad_norm": 28.375,
+      "grad_norm_var": 25.145768229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2736,
+      "loss/crossentropy": 2.0158716015517713,
+      "loss/hidden": 3.4203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19325515236705543,
+      "step": 28080
+    },
+    {
+      "epoch": 0.70225,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.1134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3422,
+      "loss/crossentropy": 2.217750224471092,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18919748738408088,
+      "step": 28090
+    },
+    {
+      "epoch": 0.7025,
+      "grad_norm": 30.0,
+      "grad_norm_var": 4.51640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3656,
+      "loss/crossentropy": 2.0768872730433943,
+      "loss/hidden": 3.348828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17947383699938654,
+      "step": 28100
+    },
+    {
+      "epoch": 0.70275,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.4052083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3558,
+      "loss/crossentropy": 2.115848197042942,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1815752826165408,
+      "step": 28110
+    },
+    {
+      "epoch": 0.703,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.878125,
+      "learning_rate": 0.0001,
+      "loss": 7.2471,
+      "loss/crossentropy": 2.0716689586639405,
+      "loss/hidden": 3.41328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19040394499897956,
+      "step": 28120
+    },
+    {
+      "epoch": 0.70325,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.8160807291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2455,
+      "loss/crossentropy": 2.171695147454739,
+      "loss/hidden": 3.3015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.186448429338634,
+      "step": 28130
+    },
+    {
+      "epoch": 0.7035,
+      "grad_norm": 30.875,
+      "grad_norm_var": 6.870833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3198,
+      "loss/crossentropy": 2.2176968052983286,
+      "loss/hidden": 3.33515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18737617712467908,
+      "step": 28140
+    },
+    {
+      "epoch": 0.70375,
+      "grad_norm": 31.5,
+      "grad_norm_var": 7.795768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3125,
+      "loss/crossentropy": 2.0401076070964335,
+      "loss/hidden": 3.305859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1749078031629324,
+      "step": 28150
+    },
+    {
+      "epoch": 0.704,
+      "grad_norm": 34.75,
+      "grad_norm_var": 5.884375,
+      "learning_rate": 0.0001,
+      "loss": 7.3385,
+      "loss/crossentropy": 2.0140421822667123,
+      "loss/hidden": 3.33046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17568598408252,
+      "step": 28160
+    },
+    {
+      "epoch": 0.70425,
+      "grad_norm": 33.25,
+      "grad_norm_var": 631.7122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2722,
+      "loss/crossentropy": 2.23368354216218,
+      "loss/hidden": 3.29140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19659436494112015,
+      "step": 28170
+    },
+    {
+      "epoch": 0.7045,
+      "grad_norm": 29.5,
+      "grad_norm_var": 5.282747395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2662,
+      "loss/crossentropy": 1.9696129187941551,
+      "loss/hidden": 3.271875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1738910754211247,
+      "step": 28180
+    },
+    {
+      "epoch": 0.70475,
+      "grad_norm": 28.25,
+      "grad_norm_var": 15.142122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2909,
+      "loss/crossentropy": 2.041120965778828,
+      "loss/hidden": 3.327734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18105307575315238,
+      "step": 28190
+    },
+    {
+      "epoch": 0.705,
+      "grad_norm": 29.0,
+      "grad_norm_var": 1.99765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2079,
+      "loss/crossentropy": 2.110129737854004,
+      "loss/hidden": 3.36328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18080106116831302,
+      "step": 28200
+    },
+    {
+      "epoch": 0.70525,
+      "grad_norm": 29.25,
+      "grad_norm_var": 1.58515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3249,
+      "loss/crossentropy": 2.148154190182686,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18392155058681964,
+      "step": 28210
+    },
+    {
+      "epoch": 0.7055,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.0155598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3179,
+      "loss/crossentropy": 2.190532484650612,
+      "loss/hidden": 3.311328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1767113889567554,
+      "step": 28220
+    },
+    {
+      "epoch": 0.70575,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.0395182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2009,
+      "loss/crossentropy": 2.175555232167244,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20121710691601039,
+      "step": 28230
+    },
+    {
+      "epoch": 0.706,
+      "grad_norm": 28.25,
+      "grad_norm_var": 2.7643229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.265,
+      "loss/crossentropy": 2.1481928557157515,
+      "loss/hidden": 3.24453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17819216307252644,
+      "step": 28240
+    },
+    {
+      "epoch": 0.70625,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.1718098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2569,
+      "loss/crossentropy": 1.997367661446333,
+      "loss/hidden": 3.343359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18763066958636046,
+      "step": 28250
+    },
+    {
+      "epoch": 0.7065,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.9580729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2066,
+      "loss/crossentropy": 2.156166338920593,
+      "loss/hidden": 3.292578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18162646796554327,
+      "step": 28260
+    },
+    {
+      "epoch": 0.70675,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.5010416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.315,
+      "loss/crossentropy": 1.9849402651190757,
+      "loss/hidden": 3.35078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18158245990052818,
+      "step": 28270
+    },
+    {
+      "epoch": 0.707,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.209375,
+      "learning_rate": 0.0001,
+      "loss": 7.2722,
+      "loss/crossentropy": 2.0745258808135985,
+      "loss/hidden": 3.318359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17213933765888215,
+      "step": 28280
+    },
+    {
+      "epoch": 0.70725,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.771875,
+      "learning_rate": 0.0001,
+      "loss": 7.2265,
+      "loss/crossentropy": 2.0611407831311226,
+      "loss/hidden": 3.2765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17584825437515975,
+      "step": 28290
+    },
+    {
+      "epoch": 0.7075,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.690625,
+      "learning_rate": 0.0001,
+      "loss": 7.3203,
+      "loss/crossentropy": 2.0343645237386228,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1779821278527379,
+      "step": 28300
+    },
+    {
+      "epoch": 0.70775,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.173958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3906,
+      "loss/crossentropy": 2.023589651286602,
+      "loss/hidden": 3.395703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19405677895992995,
+      "step": 28310
+    },
+    {
+      "epoch": 0.708,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.528059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4391,
+      "loss/crossentropy": 2.058233138918877,
+      "loss/hidden": 3.45546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18879729304462672,
+      "step": 28320
+    },
+    {
+      "epoch": 0.70825,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.81015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2291,
+      "loss/crossentropy": 1.9951315492391586,
+      "loss/hidden": 3.30234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1786661896854639,
+      "step": 28330
+    },
+    {
+      "epoch": 0.7085,
+      "grad_norm": 28.5,
+      "grad_norm_var": 3.408072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2737,
+      "loss/crossentropy": 2.1709944613277914,
+      "loss/hidden": 3.280078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18447048161178828,
+      "step": 28340
+    },
+    {
+      "epoch": 0.70875,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.0014973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3261,
+      "loss/crossentropy": 1.9754939682781696,
+      "loss/hidden": 3.368359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1782332186587155,
+      "step": 28350
+    },
+    {
+      "epoch": 0.709,
+      "grad_norm": 28.375,
+      "grad_norm_var": 3.0546223958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3341,
+      "loss/crossentropy": 1.9486119627952576,
+      "loss/hidden": 3.333203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18434599563479423,
+      "step": 28360
+    },
+    {
+      "epoch": 0.70925,
+      "grad_norm": 28.375,
+      "grad_norm_var": 3.0936848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2676,
+      "loss/crossentropy": 2.011347506940365,
+      "loss/hidden": 3.31484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17598988981917502,
+      "step": 28370
+    },
+    {
+      "epoch": 0.7095,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.278059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2683,
+      "loss/crossentropy": 2.0628366231918336,
+      "loss/hidden": 3.432421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20117868315428494,
+      "step": 28380
+    },
+    {
+      "epoch": 0.70975,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.9744140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3684,
+      "loss/crossentropy": 2.1887825660407545,
+      "loss/hidden": 3.20390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.180817243270576,
+      "step": 28390
+    },
+    {
+      "epoch": 0.71,
+      "grad_norm": 5838471168.0,
+      "grad_norm_var": 2.1304840768286884e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2715,
+      "loss/crossentropy": 1.938818733394146,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18086879644542933,
+      "step": 28400
+    },
+    {
+      "epoch": 0.71025,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.1304840755150323e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3715,
+      "loss/crossentropy": 1.953948251157999,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17758083064109087,
+      "step": 28410
+    },
+    {
+      "epoch": 0.7105,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.11640625,
+      "learning_rate": 0.0001,
+      "loss": 7.1884,
+      "loss/crossentropy": 2.3171644777059557,
+      "loss/hidden": 3.325,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19651102535426618,
+      "step": 28420
+    },
+    {
+      "epoch": 0.71075,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.7643229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3556,
+      "loss/crossentropy": 2.037036082148552,
+      "loss/hidden": 3.275,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17107345517724754,
+      "step": 28430
+    },
+    {
+      "epoch": 0.711,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.4561848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2494,
+      "loss/crossentropy": 2.127303283661604,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20090547651052476,
+      "step": 28440
+    },
+    {
+      "epoch": 0.71125,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.6809895833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3823,
+      "loss/crossentropy": 2.161235421895981,
+      "loss/hidden": 3.307421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1780310433357954,
+      "step": 28450
+    },
+    {
+      "epoch": 0.7115,
+      "grad_norm": 28.75,
+      "grad_norm_var": 1.54140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3632,
+      "loss/crossentropy": 2.1571759194135667,
+      "loss/hidden": 3.33828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19603485772386192,
+      "step": 28460
+    },
+    {
+      "epoch": 0.71175,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.364518229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2657,
+      "loss/crossentropy": 2.1602049231529237,
+      "loss/hidden": 3.325390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18435865715146066,
+      "step": 28470
+    },
+    {
+      "epoch": 0.712,
+      "grad_norm": 30.5,
+      "grad_norm_var": 10.72890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3178,
+      "loss/crossentropy": 2.001628914475441,
+      "loss/hidden": 3.3453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1918376425281167,
+      "step": 28480
+    },
+    {
+      "epoch": 0.71225,
+      "grad_norm": 67.5,
+      "grad_norm_var": 87.51712239583334,
+      "learning_rate": 0.0001,
+      "loss": 7.2839,
+      "loss/crossentropy": 1.9931135676801204,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17476971219293774,
+      "step": 28490
+    },
+    {
+      "epoch": 0.7125,
+      "grad_norm": 32.25,
+      "grad_norm_var": 96.16399739583333,
+      "learning_rate": 0.0001,
+      "loss": 7.3093,
+      "loss/crossentropy": 1.9668639115989208,
+      "loss/hidden": 3.4109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1912683444097638,
+      "step": 28500
+    },
+    {
+      "epoch": 0.71275,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.3824041782086554e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3577,
+      "loss/crossentropy": 1.9727089330554008,
+      "loss/hidden": 3.355078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17465871069580316,
+      "step": 28510
+    },
+    {
+      "epoch": 0.713,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.3824041781636367e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3469,
+      "loss/crossentropy": 2.1261328026652335,
+      "loss/hidden": 3.405859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19634310100227595,
+      "step": 28520
+    },
+    {
+      "epoch": 0.71325,
+      "grad_norm": 44.5,
+      "grad_norm_var": 14.9853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3029,
+      "loss/crossentropy": 2.029768883436918,
+      "loss/hidden": 3.30234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1729470796417445,
+      "step": 28530
+    },
+    {
+      "epoch": 0.7135,
+      "grad_norm": 30.625,
+      "grad_norm_var": 15.383072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.1945,
+      "loss/crossentropy": 2.0496911972761156,
+      "loss/hidden": 3.348828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18508521970361472,
+      "step": 28540
+    },
+    {
+      "epoch": 0.71375,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.428580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.323,
+      "loss/crossentropy": 2.064000000059605,
+      "loss/hidden": 3.3296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1856430722400546,
+      "step": 28550
+    },
+    {
+      "epoch": 0.714,
+      "grad_norm": 32.25,
+      "grad_norm_var": 5.41640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3069,
+      "loss/crossentropy": 2.0117718100547792,
+      "loss/hidden": 3.355078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17515756376087666,
+      "step": 28560
+    },
+    {
+      "epoch": 0.71425,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.8400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.1966,
+      "loss/crossentropy": 1.9726905196905136,
+      "loss/hidden": 3.383984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19007054381072522,
+      "step": 28570
+    },
+    {
+      "epoch": 0.7145,
+      "grad_norm": 28.625,
+      "grad_norm_var": 4.1416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4288,
+      "loss/crossentropy": 2.1557130038738253,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19311564825475216,
+      "step": 28580
+    },
+    {
+      "epoch": 0.71475,
+      "grad_norm": 33.5,
+      "grad_norm_var": 4.809375,
+      "learning_rate": 0.0001,
+      "loss": 7.3429,
+      "loss/crossentropy": 2.097322532534599,
+      "loss/hidden": 3.384765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18367934115231038,
+      "step": 28590
+    },
+    {
+      "epoch": 0.715,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.8895833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2557,
+      "loss/crossentropy": 2.1415718257427216,
+      "loss/hidden": 3.3859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18898510355502368,
+      "step": 28600
+    },
+    {
+      "epoch": 0.71525,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.3643229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3755,
+      "loss/crossentropy": 2.1518311858177186,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19082098407670856,
+      "step": 28610
+    },
+    {
+      "epoch": 0.7155,
+      "grad_norm": 29.25,
+      "grad_norm_var": 2.198372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3797,
+      "loss/crossentropy": 2.2079875111579894,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18495317455381155,
+      "step": 28620
+    },
+    {
+      "epoch": 0.71575,
+      "grad_norm": 27.0,
+      "grad_norm_var": 2.0608723958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2214,
+      "loss/crossentropy": 2.1074557095766067,
+      "loss/hidden": 3.30078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1770407123491168,
+      "step": 28630
+    },
+    {
+      "epoch": 0.716,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.9332682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3267,
+      "loss/crossentropy": 2.0740742847323417,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18051706086844205,
+      "step": 28640
+    },
+    {
+      "epoch": 0.71625,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.254622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1893,
+      "loss/crossentropy": 1.9990646712481976,
+      "loss/hidden": 3.221875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17936246804893016,
+      "step": 28650
+    },
+    {
+      "epoch": 0.7165,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.57890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2947,
+      "loss/crossentropy": 2.1080400705337525,
+      "loss/hidden": 3.30390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1867066564038396,
+      "step": 28660
+    },
+    {
+      "epoch": 0.71675,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.9247395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2661,
+      "loss/crossentropy": 2.074036268889904,
+      "loss/hidden": 3.356640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1867619127035141,
+      "step": 28670
+    },
+    {
+      "epoch": 0.717,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.0434895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.293,
+      "loss/crossentropy": 2.0304371282458304,
+      "loss/hidden": 3.324609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18383819735608994,
+      "step": 28680
+    },
+    {
+      "epoch": 0.71725,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.9619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4196,
+      "loss/crossentropy": 2.0870690762996675,
+      "loss/hidden": 3.4125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18118756357580423,
+      "step": 28690
+    },
+    {
+      "epoch": 0.7175,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.2504557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2668,
+      "loss/crossentropy": 2.1232191398739815,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17828409411013127,
+      "step": 28700
+    },
+    {
+      "epoch": 0.71775,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.216666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3202,
+      "loss/crossentropy": 2.093307775259018,
+      "loss/hidden": 3.308203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17778347730636596,
+      "step": 28710
+    },
+    {
+      "epoch": 0.718,
+      "grad_norm": 27.75,
+      "grad_norm_var": 2.728125,
+      "learning_rate": 0.0001,
+      "loss": 7.2286,
+      "loss/crossentropy": 2.086037275195122,
+      "loss/hidden": 3.234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17649652771651744,
+      "step": 28720
+    },
+    {
+      "epoch": 0.71825,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.872330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2935,
+      "loss/crossentropy": 2.188430291414261,
+      "loss/hidden": 3.285546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18066751547157764,
+      "step": 28730
+    },
+    {
+      "epoch": 0.7185,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.159375,
+      "learning_rate": 0.0001,
+      "loss": 7.3229,
+      "loss/crossentropy": 2.103739258646965,
+      "loss/hidden": 3.336328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18136450834572315,
+      "step": 28740
+    },
+    {
+      "epoch": 0.71875,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.215559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2833,
+      "loss/crossentropy": 1.9982567869126797,
+      "loss/hidden": 3.362890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18330774297937752,
+      "step": 28750
+    },
+    {
+      "epoch": 0.719,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.0747395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.251,
+      "loss/crossentropy": 2.1402099445462226,
+      "loss/hidden": 3.32734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1817182382568717,
+      "step": 28760
+    },
+    {
+      "epoch": 0.71925,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.667122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2284,
+      "loss/crossentropy": 1.96514939814806,
+      "loss/hidden": 3.4609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18674888461828232,
+      "step": 28770
+    },
+    {
+      "epoch": 0.7195,
+      "grad_norm": 30.875,
+      "grad_norm_var": 13.762955729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4611,
+      "loss/crossentropy": 2.109529510140419,
+      "loss/hidden": 3.3953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18532418198883532,
+      "step": 28780
+    },
+    {
+      "epoch": 0.71975,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.018684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3621,
+      "loss/crossentropy": 2.164661727845669,
+      "loss/hidden": 3.430078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22424146672710776,
+      "step": 28790
+    },
+    {
+      "epoch": 0.72,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.675,
+      "learning_rate": 0.0001,
+      "loss": 7.3163,
+      "loss/crossentropy": 2.0529872700572014,
+      "loss/hidden": 3.283203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17430850807577372,
+      "step": 28800
+    },
+    {
+      "epoch": 0.72025,
+      "grad_norm": 28.75,
+      "grad_norm_var": 3.463997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3611,
+      "loss/crossentropy": 2.091229538619518,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18271805476397276,
+      "step": 28810
+    },
+    {
+      "epoch": 0.7205,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.8306640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3029,
+      "loss/crossentropy": 1.8339010991156102,
+      "loss/hidden": 3.455078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1796142913866788,
+      "step": 28820
+    },
+    {
+      "epoch": 0.72075,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.940559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2718,
+      "loss/crossentropy": 2.2519613206386566,
+      "loss/hidden": 3.249609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1793349577113986,
+      "step": 28830
+    },
+    {
+      "epoch": 0.721,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.74140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2763,
+      "loss/crossentropy": 2.0135060526430606,
+      "loss/hidden": 3.3390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1750583214685321,
+      "step": 28840
+    },
+    {
+      "epoch": 0.72125,
+      "grad_norm": 28.375,
+      "grad_norm_var": 3.37265625,
+      "learning_rate": 0.0001,
+      "loss": 7.328,
+      "loss/crossentropy": 2.1612493455410005,
+      "loss/hidden": 3.40625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1920415310189128,
+      "step": 28850
+    },
+    {
+      "epoch": 0.7215,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.76640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3347,
+      "loss/crossentropy": 2.20034881234169,
+      "loss/hidden": 3.253515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18493964169174432,
+      "step": 28860
+    },
+    {
+      "epoch": 0.72175,
+      "grad_norm": 30.25,
+      "grad_norm_var": 0.9905598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2071,
+      "loss/crossentropy": 2.147878049314022,
+      "loss/hidden": 3.33359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18159984424710274,
+      "step": 28870
+    },
+    {
+      "epoch": 0.722,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.5624348958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2372,
+      "loss/crossentropy": 1.9990291111171246,
+      "loss/hidden": 3.358203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1772604424506426,
+      "step": 28880
+    },
+    {
+      "epoch": 0.72225,
+      "grad_norm": 28.75,
+      "grad_norm_var": 1.8525390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2288,
+      "loss/crossentropy": 2.143052602559328,
+      "loss/hidden": 3.315234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1891188271343708,
+      "step": 28890
+    },
+    {
+      "epoch": 0.7225,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.8936848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1928,
+      "loss/crossentropy": 1.9795674622058868,
+      "loss/hidden": 3.433984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17800430227071046,
+      "step": 28900
+    },
+    {
+      "epoch": 0.72275,
+      "grad_norm": 42.0,
+      "grad_norm_var": 10.046809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2488,
+      "loss/crossentropy": 2.1416710793972014,
+      "loss/hidden": 3.40625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19329405389726162,
+      "step": 28910
+    },
+    {
+      "epoch": 0.723,
+      "grad_norm": 31.25,
+      "grad_norm_var": 10.276822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2998,
+      "loss/crossentropy": 1.9632095940411092,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1833349185064435,
+      "step": 28920
+    },
+    {
+      "epoch": 0.72325,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.0395182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3208,
+      "loss/crossentropy": 2.1186553999781608,
+      "loss/hidden": 3.301171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18248404301702975,
+      "step": 28930
+    },
+    {
+      "epoch": 0.7235,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.1434895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.397,
+      "loss/crossentropy": 2.1514319077134134,
+      "loss/hidden": 3.4109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20566836930811405,
+      "step": 28940
+    },
+    {
+      "epoch": 0.72375,
+      "grad_norm": 28.25,
+      "grad_norm_var": 3.1030598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2651,
+      "loss/crossentropy": 2.0852475076913835,
+      "loss/hidden": 3.271484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17877694573253394,
+      "step": 28950
+    },
+    {
+      "epoch": 0.724,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.2171223958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3286,
+      "loss/crossentropy": 2.1540793150663378,
+      "loss/hidden": 3.395703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17957639656960964,
+      "step": 28960
+    },
+    {
+      "epoch": 0.72425,
+      "grad_norm": 28.625,
+      "grad_norm_var": 4.618489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3035,
+      "loss/crossentropy": 2.13916500210762,
+      "loss/hidden": 3.4109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2003261223435402,
+      "step": 28970
+    },
+    {
+      "epoch": 0.7245,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.435416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3627,
+      "loss/crossentropy": 2.0126554384827613,
+      "loss/hidden": 3.478515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.192663181014359,
+      "step": 28980
+    },
+    {
+      "epoch": 0.72475,
+      "grad_norm": 27.5,
+      "grad_norm_var": 12.14765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2051,
+      "loss/crossentropy": 1.9036022566258908,
+      "loss/hidden": 3.3734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1818926576524973,
+      "step": 28990
+    },
+    {
+      "epoch": 0.725,
+      "grad_norm": 28.25,
+      "grad_norm_var": 12.360416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.247,
+      "loss/crossentropy": 2.158479106426239,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2065505689010024,
+      "step": 29000
+    },
+    {
+      "epoch": 0.72525,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.158072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3395,
+      "loss/crossentropy": 2.105581759661436,
+      "loss/hidden": 3.403515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19470038171857595,
+      "step": 29010
+    },
+    {
+      "epoch": 0.7255,
+      "grad_norm": 36.75,
+      "grad_norm_var": 5.5462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3939,
+      "loss/crossentropy": 1.9542883321642877,
+      "loss/hidden": 3.4453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1957132790237665,
+      "step": 29020
+    },
+    {
+      "epoch": 0.72575,
+      "grad_norm": 33.0,
+      "grad_norm_var": 7.20390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2865,
+      "loss/crossentropy": 2.1816267639398577,
+      "loss/hidden": 3.390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2034391313791275,
+      "step": 29030
+    },
+    {
+      "epoch": 0.726,
+      "grad_norm": 30.125,
+      "grad_norm_var": 11.633072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2807,
+      "loss/crossentropy": 2.1128641098737715,
+      "loss/hidden": 3.312890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17710135243833064,
+      "step": 29040
+    },
+    {
+      "epoch": 0.72625,
+      "grad_norm": 29.75,
+      "grad_norm_var": 11.9150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2213,
+      "loss/crossentropy": 2.154743219912052,
+      "loss/hidden": 3.271875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1751030594110489,
+      "step": 29050
+    },
+    {
+      "epoch": 0.7265,
+      "grad_norm": 39.75,
+      "grad_norm_var": 63.35305989583333,
+      "learning_rate": 0.0001,
+      "loss": 7.2911,
+      "loss/crossentropy": 1.957652424275875,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1776497988961637,
+      "step": 29060
+    },
+    {
+      "epoch": 0.72675,
+      "grad_norm": 29.0,
+      "grad_norm_var": 17.909309895833335,
+      "learning_rate": 0.0001,
+      "loss": 7.2493,
+      "loss/crossentropy": 1.99792113378644,
+      "loss/hidden": 3.39140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18379787560552358,
+      "step": 29070
+    },
+    {
+      "epoch": 0.727,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.6488932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2821,
+      "loss/crossentropy": 2.0582117259502413,
+      "loss/hidden": 3.38125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18036038037389518,
+      "step": 29080
+    },
+    {
+      "epoch": 0.72725,
+      "grad_norm": 29.625,
+      "grad_norm_var": 1.8728515625,
+      "learning_rate": 0.0001,
+      "loss": 7.0774,
+      "loss/crossentropy": 2.014867861568928,
+      "loss/hidden": 3.3015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17281777542084456,
+      "step": 29090
+    },
+    {
+      "epoch": 0.7275,
+      "grad_norm": 28.625,
+      "grad_norm_var": 3.658333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3509,
+      "loss/crossentropy": 2.085656464099884,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17972098607569933,
+      "step": 29100
+    },
+    {
+      "epoch": 0.72775,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.487434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3234,
+      "loss/crossentropy": 2.0445970110595226,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16965955086052417,
+      "step": 29110
+    },
+    {
+      "epoch": 0.728,
+      "grad_norm": 28.0,
+      "grad_norm_var": 1.3020182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2674,
+      "loss/crossentropy": 2.2715115547180176,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1960804032161832,
+      "step": 29120
+    },
+    {
+      "epoch": 0.72825,
+      "grad_norm": 29.625,
+      "grad_norm_var": 5.242708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3008,
+      "loss/crossentropy": 2.1775743424892426,
+      "loss/hidden": 3.32734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18752657640725373,
+      "step": 29130
+    },
+    {
+      "epoch": 0.7285,
+      "grad_norm": 38.5,
+      "grad_norm_var": 3.132886831921837e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.219,
+      "loss/crossentropy": 1.903736773133278,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17021676823496817,
+      "step": 29140
+    },
+    {
+      "epoch": 0.72875,
+      "grad_norm": 38.75,
+      "grad_norm_var": 20.174739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3486,
+      "loss/crossentropy": 2.1836992099881174,
+      "loss/hidden": 3.478125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2182300578802824,
+      "step": 29150
+    },
+    {
+      "epoch": 0.729,
+      "grad_norm": 32.5,
+      "grad_norm_var": 17.109830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2701,
+      "loss/crossentropy": 2.0530862897634505,
+      "loss/hidden": 3.2484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18154476722702384,
+      "step": 29160
+    },
+    {
+      "epoch": 0.72925,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.2494140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2304,
+      "loss/crossentropy": 2.0158081978559492,
+      "loss/hidden": 3.440625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18534094132483006,
+      "step": 29170
+    },
+    {
+      "epoch": 0.7295,
+      "grad_norm": 30.25,
+      "grad_norm_var": 43.02682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3198,
+      "loss/crossentropy": 2.254640057682991,
+      "loss/hidden": 3.26953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19149381387978792,
+      "step": 29180
+    },
+    {
+      "epoch": 0.72975,
+      "grad_norm": 32.25,
+      "grad_norm_var": 42.57265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3284,
+      "loss/crossentropy": 2.1554953277111055,
+      "loss/hidden": 3.38125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1914082583039999,
+      "step": 29190
+    },
+    {
+      "epoch": 0.73,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.520247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.188,
+      "loss/crossentropy": 2.095432303100824,
+      "loss/hidden": 3.308203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17190228644758462,
+      "step": 29200
+    },
+    {
+      "epoch": 0.73025,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.8103515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2669,
+      "loss/crossentropy": 1.9353884100914,
+      "loss/hidden": 3.3296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1733547686599195,
+      "step": 29210
+    },
+    {
+      "epoch": 0.7305,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.64375,
+      "learning_rate": 0.0001,
+      "loss": 7.38,
+      "loss/crossentropy": 2.1763652101159097,
+      "loss/hidden": 3.349609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19759245738387107,
+      "step": 29220
+    },
+    {
+      "epoch": 0.73075,
+      "grad_norm": 29.625,
+      "grad_norm_var": 1.9811848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.316,
+      "loss/crossentropy": 2.059311767667532,
+      "loss/hidden": 3.273046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1732610635459423,
+      "step": 29230
+    },
+    {
+      "epoch": 0.731,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.6296223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2387,
+      "loss/crossentropy": 2.1624471306800843,
+      "loss/hidden": 3.27421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18443353176116944,
+      "step": 29240
+    },
+    {
+      "epoch": 0.73125,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.1622395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.255,
+      "loss/crossentropy": 2.087037294358015,
+      "loss/hidden": 3.46484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18923133062198758,
+      "step": 29250
+    },
+    {
+      "epoch": 0.7315,
+      "grad_norm": 28.625,
+      "grad_norm_var": 1.5389973958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1715,
+      "loss/crossentropy": 2.0068927347660064,
+      "loss/hidden": 3.336328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1776790694333613,
+      "step": 29260
+    },
+    {
+      "epoch": 0.73175,
+      "grad_norm": 34.75,
+      "grad_norm_var": 3.0125,
+      "learning_rate": 0.0001,
+      "loss": 7.3044,
+      "loss/crossentropy": 2.237174994498491,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1987008555792272,
+      "step": 29270
+    },
+    {
+      "epoch": 0.732,
+      "grad_norm": 32.75,
+      "grad_norm_var": 4.869791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3198,
+      "loss/crossentropy": 2.073240910470486,
+      "loss/hidden": 3.3578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19791424218565226,
+      "step": 29280
+    },
+    {
+      "epoch": 0.73225,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.595833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3329,
+      "loss/crossentropy": 2.123592960834503,
+      "loss/hidden": 3.3375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18660867251455784,
+      "step": 29290
+    },
+    {
+      "epoch": 0.7325,
+      "grad_norm": 27.25,
+      "grad_norm_var": 2.0957682291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2039,
+      "loss/crossentropy": 2.1879305720329283,
+      "loss/hidden": 3.269921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18721728827804326,
+      "step": 29300
+    },
+    {
+      "epoch": 0.73275,
+      "grad_norm": 26.375,
+      "grad_norm_var": 23.684830729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3075,
+      "loss/crossentropy": 2.0866588428616524,
+      "loss/hidden": 3.3125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17703543622046708,
+      "step": 29310
+    },
+    {
+      "epoch": 0.733,
+      "grad_norm": 31.375,
+      "grad_norm_var": 23.417708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2536,
+      "loss/crossentropy": 2.250990514457226,
+      "loss/hidden": 3.349609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18253531642258167,
+      "step": 29320
+    },
+    {
+      "epoch": 0.73325,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.414322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2226,
+      "loss/crossentropy": 2.102177432179451,
+      "loss/hidden": 3.341015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19504718706011773,
+      "step": 29330
+    },
+    {
+      "epoch": 0.7335,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.820572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1243,
+      "loss/crossentropy": 2.0512589052319528,
+      "loss/hidden": 3.383984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18182420805096627,
+      "step": 29340
+    },
+    {
+      "epoch": 0.73375,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.9934895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3111,
+      "loss/crossentropy": 2.143858629465103,
+      "loss/hidden": 3.480078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19919059351086615,
+      "step": 29350
+    },
+    {
+      "epoch": 0.734,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.80625,
+      "learning_rate": 0.0001,
+      "loss": 7.2538,
+      "loss/crossentropy": 2.0024397730827332,
+      "loss/hidden": 3.255859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17439783103764056,
+      "step": 29360
+    },
+    {
+      "epoch": 0.73425,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.5143229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.1658,
+      "loss/crossentropy": 1.8820870153605938,
+      "loss/hidden": 3.353515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17721544997766614,
+      "step": 29370
+    },
+    {
+      "epoch": 0.7345,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.33125,
+      "learning_rate": 0.0001,
+      "loss": 7.2407,
+      "loss/crossentropy": 2.068721887469292,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1961481623351574,
+      "step": 29380
+    },
+    {
+      "epoch": 0.73475,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.603125,
+      "learning_rate": 0.0001,
+      "loss": 7.252,
+      "loss/crossentropy": 2.016729524731636,
+      "loss/hidden": 3.3640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18200870957225562,
+      "step": 29390
+    },
+    {
+      "epoch": 0.735,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.48515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3778,
+      "loss/crossentropy": 2.2266832500696183,
+      "loss/hidden": 3.271875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18430891614407302,
+      "step": 29400
+    },
+    {
+      "epoch": 0.73525,
+      "grad_norm": 31.0,
+      "grad_norm_var": 7.8931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2621,
+      "loss/crossentropy": 1.9326732844114303,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19005155926570297,
+      "step": 29410
+    },
+    {
+      "epoch": 0.7355,
+      "grad_norm": 29.25,
+      "grad_norm_var": 6.9681640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3399,
+      "loss/crossentropy": 2.017210554331541,
+      "loss/hidden": 3.240625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.190559295238927,
+      "step": 29420
+    },
+    {
+      "epoch": 0.73575,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.7583333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2686,
+      "loss/crossentropy": 1.7766004413366319,
+      "loss/hidden": 3.479296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18161138612776995,
+      "step": 29430
+    },
+    {
+      "epoch": 0.736,
+      "grad_norm": 30.625,
+      "grad_norm_var": 7.8041015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4176,
+      "loss/crossentropy": 2.1919517919421194,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19109788965433835,
+      "step": 29440
+    },
+    {
+      "epoch": 0.73625,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.630989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2415,
+      "loss/crossentropy": 2.187614656984806,
+      "loss/hidden": 3.294921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17972742971032857,
+      "step": 29450
+    },
+    {
+      "epoch": 0.7365,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.4166015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2987,
+      "loss/crossentropy": 2.2710886627435682,
+      "loss/hidden": 3.292578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1898843079805374,
+      "step": 29460
+    },
+    {
+      "epoch": 0.73675,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.787955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.1919,
+      "loss/crossentropy": 2.1406801946461202,
+      "loss/hidden": 3.28671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18407909590750932,
+      "step": 29470
+    },
+    {
+      "epoch": 0.737,
+      "grad_norm": 28.875,
+      "grad_norm_var": 10.710416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3633,
+      "loss/crossentropy": 2.117516166716814,
+      "loss/hidden": 3.287109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1789137402549386,
+      "step": 29480
+    },
+    {
+      "epoch": 0.73725,
+      "grad_norm": 29.25,
+      "grad_norm_var": 16.240559895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.1497,
+      "loss/crossentropy": 2.0938372761011124,
+      "loss/hidden": 3.237109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16651670411229133,
+      "step": 29490
+    },
+    {
+      "epoch": 0.7375,
+      "grad_norm": 31.0,
+      "grad_norm_var": 8.467122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.1916,
+      "loss/crossentropy": 2.14051573574543,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17684268727898597,
+      "step": 29500
+    },
+    {
+      "epoch": 0.73775,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.834375,
+      "learning_rate": 0.0001,
+      "loss": 7.2436,
+      "loss/crossentropy": 1.9776017241179944,
+      "loss/hidden": 3.290234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17606793977320195,
+      "step": 29510
+    },
+    {
+      "epoch": 0.738,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.5052083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2794,
+      "loss/crossentropy": 2.1177644938230515,
+      "loss/hidden": 3.40078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19227085784077644,
+      "step": 29520
+    },
+    {
+      "epoch": 0.73825,
+      "grad_norm": 34.75,
+      "grad_norm_var": 6.0197265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2809,
+      "loss/crossentropy": 2.0195805758237837,
+      "loss/hidden": 3.32421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16698714587837457,
+      "step": 29530
+    },
+    {
+      "epoch": 0.7385,
+      "grad_norm": 31.25,
+      "grad_norm_var": 5.507747395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2993,
+      "loss/crossentropy": 2.17318941950798,
+      "loss/hidden": 3.315234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1887758129276335,
+      "step": 29540
+    },
+    {
+      "epoch": 0.73875,
+      "grad_norm": 28.875,
+      "grad_norm_var": 1.6400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2153,
+      "loss/crossentropy": 2.0143275789916517,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16856615301221609,
+      "step": 29550
+    },
+    {
+      "epoch": 0.739,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.662239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3949,
+      "loss/crossentropy": 2.2027142092585565,
+      "loss/hidden": 3.379296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18870100285857916,
+      "step": 29560
+    },
+    {
+      "epoch": 0.73925,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.7436848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3728,
+      "loss/crossentropy": 1.9897805616259574,
+      "loss/hidden": 3.44609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21316022276878357,
+      "step": 29570
+    },
+    {
+      "epoch": 0.7395,
+      "grad_norm": 36.0,
+      "grad_norm_var": 3.5747395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2612,
+      "loss/crossentropy": 2.169425293803215,
+      "loss/hidden": 3.357421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18564392458647488,
+      "step": 29580
+    },
+    {
+      "epoch": 0.73975,
+      "grad_norm": 33.5,
+      "grad_norm_var": 5.126497395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.1713,
+      "loss/crossentropy": 1.8968232870101929,
+      "loss/hidden": 3.403125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17540717963129282,
+      "step": 29590
+    },
+    {
+      "epoch": 0.74,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.905989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3553,
+      "loss/crossentropy": 2.120753511786461,
+      "loss/hidden": 3.398046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1904517425224185,
+      "step": 29600
+    },
+    {
+      "epoch": 0.74025,
+      "grad_norm": 30.375,
+      "grad_norm_var": 4.1375,
+      "learning_rate": 0.0001,
+      "loss": 7.3046,
+      "loss/crossentropy": 2.1590781211853027,
+      "loss/hidden": 3.3734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19008232783526183,
+      "step": 29610
+    },
+    {
+      "epoch": 0.7405,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.600455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3147,
+      "loss/crossentropy": 1.9427101552486419,
+      "loss/hidden": 3.308984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.170235357247293,
+      "step": 29620
+    },
+    {
+      "epoch": 0.74075,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.8582682291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3301,
+      "loss/crossentropy": 2.0426175825297834,
+      "loss/hidden": 3.336328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1734724512323737,
+      "step": 29630
+    },
+    {
+      "epoch": 0.741,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.0462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3503,
+      "loss/crossentropy": 2.1062657207250597,
+      "loss/hidden": 3.330078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1793691584840417,
+      "step": 29640
+    },
+    {
+      "epoch": 0.74125,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.2684895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2387,
+      "loss/crossentropy": 2.0198217228055,
+      "loss/hidden": 3.3234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17088997922837734,
+      "step": 29650
+    },
+    {
+      "epoch": 0.7415,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.3677083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2345,
+      "loss/crossentropy": 2.131781388819218,
+      "loss/hidden": 3.33828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18850240260362625,
+      "step": 29660
+    },
+    {
+      "epoch": 0.74175,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.145247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2585,
+      "loss/crossentropy": 2.1001341193914413,
+      "loss/hidden": 3.291015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18017663452774285,
+      "step": 29670
+    },
+    {
+      "epoch": 0.742,
+      "grad_norm": 30.375,
+      "grad_norm_var": 4.264322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3288,
+      "loss/crossentropy": 2.1850179612636564,
+      "loss/hidden": 3.337890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1874097477644682,
+      "step": 29680
+    },
+    {
+      "epoch": 0.74225,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.655143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.284,
+      "loss/crossentropy": 2.088411360234022,
+      "loss/hidden": 3.348828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1895237101241946,
+      "step": 29690
+    },
+    {
+      "epoch": 0.7425,
+      "grad_norm": 28.125,
+      "grad_norm_var": 3.039322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1583,
+      "loss/crossentropy": 2.0124595299363137,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17138318605720998,
+      "step": 29700
+    },
+    {
+      "epoch": 0.74275,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.5801432291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2585,
+      "loss/crossentropy": 2.1081888318061828,
+      "loss/hidden": 3.287109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17212112471461297,
+      "step": 29710
+    },
+    {
+      "epoch": 0.743,
+      "grad_norm": 30.25,
+      "grad_norm_var": 4.846809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2807,
+      "loss/crossentropy": 2.052864673733711,
+      "loss/hidden": 3.30546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17533118948340415,
+      "step": 29720
+    },
+    {
+      "epoch": 0.74325,
+      "grad_norm": 28.875,
+      "grad_norm_var": 1.8083333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2476,
+      "loss/crossentropy": 2.0982922323048117,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18855995871126652,
+      "step": 29730
+    },
+    {
+      "epoch": 0.7435,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.3103515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3216,
+      "loss/crossentropy": 2.0989072144031526,
+      "loss/hidden": 3.290625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1823441507294774,
+      "step": 29740
+    },
+    {
+      "epoch": 0.74375,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.7979166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2959,
+      "loss/crossentropy": 2.1087387457489966,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19064594879746438,
+      "step": 29750
+    },
+    {
+      "epoch": 0.744,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.8525390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4191,
+      "loss/crossentropy": 1.9283446200191974,
+      "loss/hidden": 3.3984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1755460049957037,
+      "step": 29760
+    },
+    {
+      "epoch": 0.74425,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.370247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1647,
+      "loss/crossentropy": 2.0763636574149134,
+      "loss/hidden": 3.30078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18310361802577974,
+      "step": 29770
+    },
+    {
+      "epoch": 0.7445,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.1337890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3702,
+      "loss/crossentropy": 2.1763921469449996,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18296955525875092,
+      "step": 29780
+    },
+    {
+      "epoch": 0.74475,
+      "grad_norm": 35.0,
+      "grad_norm_var": 4.333072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2519,
+      "loss/crossentropy": 2.101426640152931,
+      "loss/hidden": 3.403125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18623395673930646,
+      "step": 29790
+    },
+    {
+      "epoch": 0.745,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.4559895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2323,
+      "loss/crossentropy": 2.1355843022465706,
+      "loss/hidden": 3.42734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19213641248643398,
+      "step": 29800
+    },
+    {
+      "epoch": 0.74525,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.5947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3239,
+      "loss/crossentropy": 2.138861904293299,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19748148396611215,
+      "step": 29810
+    },
+    {
+      "epoch": 0.7455,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.6455729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2485,
+      "loss/crossentropy": 2.074553743004799,
+      "loss/hidden": 3.25390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17721196599304676,
+      "step": 29820
+    },
+    {
+      "epoch": 0.74575,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.919205729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2774,
+      "loss/crossentropy": 2.206491988897324,
+      "loss/hidden": 3.413671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19982497543096542,
+      "step": 29830
+    },
+    {
+      "epoch": 0.746,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.2625,
+      "learning_rate": 0.0001,
+      "loss": 7.1951,
+      "loss/crossentropy": 2.032223401963711,
+      "loss/hidden": 3.3328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18081800378859042,
+      "step": 29840
+    },
+    {
+      "epoch": 0.74625,
+      "grad_norm": 28.125,
+      "grad_norm_var": 2.6455729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2018,
+      "loss/crossentropy": 2.1063923329114913,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17570548821240664,
+      "step": 29850
+    },
+    {
+      "epoch": 0.7465,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.1328868320619617e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2075,
+      "loss/crossentropy": 2.008588743954897,
+      "loss/hidden": 3.26640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16772094201296567,
+      "step": 29860
+    },
+    {
+      "epoch": 0.74675,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.1328868314940877e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2256,
+      "loss/crossentropy": 2.327219396829605,
+      "loss/hidden": 3.30078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1916476283222437,
+      "step": 29870
+    },
+    {
+      "epoch": 0.747,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.06015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2006,
+      "loss/crossentropy": 2.1967759788036347,
+      "loss/hidden": 3.255859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.181925462000072,
+      "step": 29880
+    },
+    {
+      "epoch": 0.74725,
+      "grad_norm": 29.0,
+      "grad_norm_var": 1.07265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3302,
+      "loss/crossentropy": 2.0784429952502252,
+      "loss/hidden": 3.4046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20122398622334003,
+      "step": 29890
+    },
+    {
+      "epoch": 0.7475,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.16015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2241,
+      "loss/crossentropy": 2.044815068691969,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17522517647594213,
+      "step": 29900
+    },
+    {
+      "epoch": 0.74775,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.7848307291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.241,
+      "loss/crossentropy": 2.0202932924032213,
+      "loss/hidden": 3.4359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18710787147283553,
+      "step": 29910
+    },
+    {
+      "epoch": 0.748,
+      "grad_norm": 35.75,
+      "grad_norm_var": 4.892643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4702,
+      "loss/crossentropy": 1.9623441621661186,
+      "loss/hidden": 3.460546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18277720119804144,
+      "step": 29920
+    },
+    {
+      "epoch": 0.74825,
+      "grad_norm": 33.5,
+      "grad_norm_var": 32.8087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2328,
+      "loss/crossentropy": 2.1580356508493423,
+      "loss/hidden": 3.273828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18409592770040034,
+      "step": 29930
+    },
+    {
+      "epoch": 0.7485,
+      "grad_norm": 29.5,
+      "grad_norm_var": 5.335416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2307,
+      "loss/crossentropy": 1.9020950391888618,
+      "loss/hidden": 3.2171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16977879907935858,
+      "step": 29940
+    },
+    {
+      "epoch": 0.74875,
+      "grad_norm": 28.5,
+      "grad_norm_var": 3.9124348958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.1602,
+      "loss/crossentropy": 1.9830825686454774,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19025738779455423,
+      "step": 29950
+    },
+    {
+      "epoch": 0.749,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.4872395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2886,
+      "loss/crossentropy": 2.0990236908197404,
+      "loss/hidden": 3.404296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20246837567538023,
+      "step": 29960
+    },
+    {
+      "epoch": 0.74925,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.5291015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3944,
+      "loss/crossentropy": 2.0726598769426348,
+      "loss/hidden": 3.27890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17855302877724172,
+      "step": 29970
+    },
+    {
+      "epoch": 0.7495,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.155989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.0988,
+      "loss/crossentropy": 2.0020054180175064,
+      "loss/hidden": 3.27890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1707400310318917,
+      "step": 29980
+    },
+    {
+      "epoch": 0.74975,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.613541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3437,
+      "loss/crossentropy": 2.1497740030288695,
+      "loss/hidden": 3.306640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18407329265028238,
+      "step": 29990
+    },
+    {
+      "epoch": 0.75,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.3348307291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3011,
+      "loss/crossentropy": 2.10102431550622,
+      "loss/hidden": 3.404296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18685767855495214,
+      "step": 30000
+    },
+    {
+      "epoch": 0.75025,
+      "grad_norm": 28.75,
+      "grad_norm_var": 1.8431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2246,
+      "loss/crossentropy": 2.2319724306464197,
+      "loss/hidden": 3.389453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21383447982370854,
+      "step": 30010
+    },
+    {
+      "epoch": 0.7505,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.6244140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3341,
+      "loss/crossentropy": 2.021278867125511,
+      "loss/hidden": 3.43359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20931231547147036,
+      "step": 30020
+    },
+    {
+      "epoch": 0.75075,
+      "grad_norm": 29.375,
+      "grad_norm_var": 4.070572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2765,
+      "loss/crossentropy": 2.1085473522543907,
+      "loss/hidden": 3.330078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18091097762808203,
+      "step": 30030
+    },
+    {
+      "epoch": 0.751,
+      "grad_norm": 31.75,
+      "grad_norm_var": 4.53125,
+      "learning_rate": 0.0001,
+      "loss": 7.3646,
+      "loss/crossentropy": 2.1863505825400353,
+      "loss/hidden": 3.35078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1947094239294529,
+      "step": 30040
+    },
+    {
+      "epoch": 0.75125,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.1030598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2667,
+      "loss/crossentropy": 1.9493524998426437,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18073509354144335,
+      "step": 30050
+    },
+    {
+      "epoch": 0.7515,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.85625,
+      "learning_rate": 0.0001,
+      "loss": 7.3301,
+      "loss/crossentropy": 2.0162834845483304,
+      "loss/hidden": 3.274609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17740605510771273,
+      "step": 30060
+    },
+    {
+      "epoch": 0.75175,
+      "grad_norm": 30.0,
+      "grad_norm_var": 1.6854166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3117,
+      "loss/crossentropy": 2.247773203253746,
+      "loss/hidden": 3.307421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1955332661047578,
+      "step": 30070
+    },
+    {
+      "epoch": 0.752,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.721809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2713,
+      "loss/crossentropy": 1.8765224024653435,
+      "loss/hidden": 3.3640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18920745309442283,
+      "step": 30080
+    },
+    {
+      "epoch": 0.75225,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.4228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2952,
+      "loss/crossentropy": 2.0629479378461837,
+      "loss/hidden": 3.31328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1822500554844737,
+      "step": 30090
+    },
+    {
+      "epoch": 0.7525,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.8885416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2476,
+      "loss/crossentropy": 2.0535168081521986,
+      "loss/hidden": 3.379296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1875293802469969,
+      "step": 30100
+    },
+    {
+      "epoch": 0.75275,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.8353515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3444,
+      "loss/crossentropy": 2.153412875533104,
+      "loss/hidden": 3.379296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20092597529292106,
+      "step": 30110
+    },
+    {
+      "epoch": 0.753,
+      "grad_norm": 29.75,
+      "grad_norm_var": 13.36015625,
+      "learning_rate": 0.0001,
+      "loss": 7.1798,
+      "loss/crossentropy": 2.135878336429596,
+      "loss/hidden": 3.26328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17593010412529111,
+      "step": 30120
+    },
+    {
+      "epoch": 0.75325,
+      "grad_norm": 33.0,
+      "grad_norm_var": 18.307747395833335,
+      "learning_rate": 0.0001,
+      "loss": 7.2136,
+      "loss/crossentropy": 2.1997751981019973,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18918296322226524,
+      "step": 30130
+    },
+    {
+      "epoch": 0.7535,
+      "grad_norm": 39.5,
+      "grad_norm_var": 3.0444333210866637e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.332,
+      "loss/crossentropy": 1.9713260725140571,
+      "loss/hidden": 3.4015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19746107552200556,
+      "step": 30140
+    },
+    {
+      "epoch": 0.75375,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.04443332031603e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3182,
+      "loss/crossentropy": 2.0566159434616567,
+      "loss/hidden": 3.31015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17785473773255944,
+      "step": 30150
+    },
+    {
+      "epoch": 0.754,
+      "grad_norm": 30.75,
+      "grad_norm_var": 4.223372395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2001,
+      "loss/crossentropy": 1.8348468966782092,
+      "loss/hidden": 3.287109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.15625384994782507,
+      "step": 30160
+    },
+    {
+      "epoch": 0.75425,
+      "grad_norm": 35.25,
+      "grad_norm_var": 4.393489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3332,
+      "loss/crossentropy": 1.963544089347124,
+      "loss/hidden": 3.5234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22075258214026688,
+      "step": 30170
+    },
+    {
+      "epoch": 0.7545,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.516080729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2782,
+      "loss/crossentropy": 1.9644573964178562,
+      "loss/hidden": 3.4171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18735937895253302,
+      "step": 30180
+    },
+    {
+      "epoch": 0.75475,
+      "grad_norm": 29.5,
+      "grad_norm_var": 4.713541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2855,
+      "loss/crossentropy": 2.0815071538090706,
+      "loss/hidden": 3.3453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18255099412053824,
+      "step": 30190
+    },
+    {
+      "epoch": 0.755,
+      "grad_norm": 29.375,
+      "grad_norm_var": 7.876822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2586,
+      "loss/crossentropy": 2.0835547246038915,
+      "loss/hidden": 3.294921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18153475932776927,
+      "step": 30200
+    },
+    {
+      "epoch": 0.75525,
+      "grad_norm": 44.0,
+      "grad_norm_var": 13.9884765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3384,
+      "loss/crossentropy": 2.1766451790928842,
+      "loss/hidden": 3.412890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1941084787249565,
+      "step": 30210
+    },
+    {
+      "epoch": 0.7555,
+      "grad_norm": 28.125,
+      "grad_norm_var": 16.192122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3195,
+      "loss/crossentropy": 2.1572922460734842,
+      "loss/hidden": 3.2859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18345598923042417,
+      "step": 30220
+    },
+    {
+      "epoch": 0.75575,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.8363932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1101,
+      "loss/crossentropy": 1.9785410642623902,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1788899280130863,
+      "step": 30230
+    },
+    {
+      "epoch": 0.756,
+      "grad_norm": 30.75,
+      "grad_norm_var": 14.198958333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2872,
+      "loss/crossentropy": 2.032851605117321,
+      "loss/hidden": 3.440625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19580046012997626,
+      "step": 30240
+    },
+    {
+      "epoch": 0.75625,
+      "grad_norm": 31.375,
+      "grad_norm_var": 17.10390625,
+      "learning_rate": 0.0001,
+      "loss": 7.112,
+      "loss/crossentropy": 2.0762212831526994,
+      "loss/hidden": 3.31328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1714485607109964,
+      "step": 30250
+    },
+    {
+      "epoch": 0.7565,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.03125,
+      "learning_rate": 0.0001,
+      "loss": 7.2301,
+      "loss/crossentropy": 2.2272632330656053,
+      "loss/hidden": 3.43046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20788154676556586,
+      "step": 30260
+    },
+    {
+      "epoch": 0.75675,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.09140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2818,
+      "loss/crossentropy": 2.0946265175938605,
+      "loss/hidden": 3.28203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18614709600806237,
+      "step": 30270
+    },
+    {
+      "epoch": 0.757,
+      "grad_norm": 28.375,
+      "grad_norm_var": 0.8514973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1706,
+      "loss/crossentropy": 2.0933008804917335,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17898790668696166,
+      "step": 30280
+    },
+    {
+      "epoch": 0.75725,
+      "grad_norm": 27.875,
+      "grad_norm_var": 1.0989583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3146,
+      "loss/crossentropy": 2.122311710566282,
+      "loss/hidden": 3.312890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16947407890111207,
+      "step": 30290
+    },
+    {
+      "epoch": 0.7575,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.0152302917171804e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3609,
+      "loss/crossentropy": 2.206487476825714,
+      "loss/hidden": 3.308203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19262611195445062,
+      "step": 30300
+    },
+    {
+      "epoch": 0.75775,
+      "grad_norm": 27.625,
+      "grad_norm_var": 3.122330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.208,
+      "loss/crossentropy": 2.166238710284233,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19571597315371037,
+      "step": 30310
+    },
+    {
+      "epoch": 0.758,
+      "grad_norm": 27.75,
+      "grad_norm_var": 2.3692057291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3066,
+      "loss/crossentropy": 2.1176643908023833,
+      "loss/hidden": 3.2671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17054121531546115,
+      "step": 30320
+    },
+    {
+      "epoch": 0.75825,
+      "grad_norm": 34.0,
+      "grad_norm_var": 238.5572265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2952,
+      "loss/crossentropy": 1.9073448725044728,
+      "loss/hidden": 3.3484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1762677811086178,
+      "step": 30330
+    },
+    {
+      "epoch": 0.7585,
+      "grad_norm": 28.5,
+      "grad_norm_var": 231.375,
+      "learning_rate": 0.0001,
+      "loss": 7.1593,
+      "loss/crossentropy": 2.0104400724172593,
+      "loss/hidden": 3.259765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17894950192421674,
+      "step": 30340
+    },
+    {
+      "epoch": 0.75875,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.9681640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2992,
+      "loss/crossentropy": 2.0130166873335837,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17820241563022138,
+      "step": 30350
+    },
+    {
+      "epoch": 0.759,
+      "grad_norm": 34.75,
+      "grad_norm_var": 3.71640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2691,
+      "loss/crossentropy": 2.0536349333822725,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18395173139870166,
+      "step": 30360
+    },
+    {
+      "epoch": 0.75925,
+      "grad_norm": 29.875,
+      "grad_norm_var": 25.2056640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2065,
+      "loss/crossentropy": 2.1257813230156897,
+      "loss/hidden": 3.39453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17920450903475285,
+      "step": 30370
+    },
+    {
+      "epoch": 0.7595,
+      "grad_norm": 31.125,
+      "grad_norm_var": 24.911393229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2013,
+      "loss/crossentropy": 2.2414271771907806,
+      "loss/hidden": 3.2984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19076382871717215,
+      "step": 30380
+    },
+    {
+      "epoch": 0.75975,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.603125,
+      "learning_rate": 0.0001,
+      "loss": 7.163,
+      "loss/crossentropy": 2.077455496788025,
+      "loss/hidden": 3.286328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.189267154969275,
+      "step": 30390
+    },
+    {
+      "epoch": 0.76,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.5791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3401,
+      "loss/crossentropy": 2.0604166358709337,
+      "loss/hidden": 3.347265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1891783157363534,
+      "step": 30400
+    },
+    {
+      "epoch": 0.76025,
+      "grad_norm": 28.5,
+      "grad_norm_var": 4.080208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2248,
+      "loss/crossentropy": 2.021010288596153,
+      "loss/hidden": 3.323046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1886049335822463,
+      "step": 30410
+    },
+    {
+      "epoch": 0.7605,
+      "grad_norm": 34.75,
+      "grad_norm_var": 2.6080729166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2918,
+      "loss/crossentropy": 2.104072753340006,
+      "loss/hidden": 3.34140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18789286576211453,
+      "step": 30420
+    },
+    {
+      "epoch": 0.76075,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.983268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3593,
+      "loss/crossentropy": 2.077750712633133,
+      "loss/hidden": 3.368359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18233956769108772,
+      "step": 30430
+    },
+    {
+      "epoch": 0.761,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.0737770858689853e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3333,
+      "loss/crossentropy": 2.1295972421765326,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18910994809120893,
+      "step": 30440
+    },
+    {
+      "epoch": 0.76125,
+      "grad_norm": 32.25,
+      "grad_norm_var": 27.800455729166668,
+      "learning_rate": 0.0001,
+      "loss": 7.4099,
+      "loss/crossentropy": 2.0139816865324973,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18354472760111093,
+      "step": 30450
+    },
+    {
+      "epoch": 0.7615,
+      "grad_norm": 29.0,
+      "grad_norm_var": 164.29837239583333,
+      "learning_rate": 0.0001,
+      "loss": 7.1649,
+      "loss/crossentropy": 2.0612717509269713,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1756343010812998,
+      "step": 30460
+    },
+    {
+      "epoch": 0.76175,
+      "grad_norm": 28.0,
+      "grad_norm_var": 167.95104166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2015,
+      "loss/crossentropy": 2.110891255736351,
+      "loss/hidden": 3.349609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19027273952960969,
+      "step": 30470
+    },
+    {
+      "epoch": 0.762,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.1747395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.277,
+      "loss/crossentropy": 2.1438290625810623,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.202046955563128,
+      "step": 30480
+    },
+    {
+      "epoch": 0.76225,
+      "grad_norm": 28.5,
+      "grad_norm_var": 22.799739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2273,
+      "loss/crossentropy": 1.9993860483169557,
+      "loss/hidden": 3.394140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18927458357065916,
+      "step": 30490
+    },
+    {
+      "epoch": 0.7625,
+      "grad_norm": 28.75,
+      "grad_norm_var": 27.1634765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2857,
+      "loss/crossentropy": 2.0218796797096728,
+      "loss/hidden": 3.38984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19677631575614213,
+      "step": 30500
+    },
+    {
+      "epoch": 0.76275,
+      "grad_norm": 30.0,
+      "grad_norm_var": 31.164583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2166,
+      "loss/crossentropy": 2.013028331845999,
+      "loss/hidden": 3.326171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18638179991394282,
+      "step": 30510
+    },
+    {
+      "epoch": 0.763,
+      "grad_norm": 30.875,
+      "grad_norm_var": 15.534309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2954,
+      "loss/crossentropy": 2.013714115321636,
+      "loss/hidden": 3.4078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1986254785209894,
+      "step": 30520
+    },
+    {
+      "epoch": 0.76325,
+      "grad_norm": 36.75,
+      "grad_norm_var": 13.345572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2507,
+      "loss/crossentropy": 2.0372176095843315,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18047925494611264,
+      "step": 30530
+    },
+    {
+      "epoch": 0.7635,
+      "grad_norm": 29.5,
+      "grad_norm_var": 8.724739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1431,
+      "loss/crossentropy": 1.8315919287502767,
+      "loss/hidden": 3.298828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16889077816158532,
+      "step": 30540
+    },
+    {
+      "epoch": 0.76375,
+      "grad_norm": 30.0,
+      "grad_norm_var": 6.551822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3346,
+      "loss/crossentropy": 2.0934281766414644,
+      "loss/hidden": 3.321484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18256967328488827,
+      "step": 30550
+    },
+    {
+      "epoch": 0.764,
+      "grad_norm": 35.5,
+      "grad_norm_var": 10.449739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1819,
+      "loss/crossentropy": 1.9229003690183162,
+      "loss/hidden": 3.341015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17857125196605922,
+      "step": 30560
+    },
+    {
+      "epoch": 0.76425,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.8202487216409324e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2603,
+      "loss/crossentropy": 1.8773075222969056,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.182831759005785,
+      "step": 30570
+    },
+    {
+      "epoch": 0.7645,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.820248721478054e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2925,
+      "loss/crossentropy": 2.222408211231232,
+      "loss/hidden": 3.37109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1848822958767414,
+      "step": 30580
+    },
+    {
+      "epoch": 0.76475,
+      "grad_norm": 30.5,
+      "grad_norm_var": 6.410872395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.1576,
+      "loss/crossentropy": 2.0349696803838015,
+      "loss/hidden": 3.259375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17177348136901854,
+      "step": 30590
+    },
+    {
+      "epoch": 0.765,
+      "grad_norm": 37.5,
+      "grad_norm_var": 7.037239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2271,
+      "loss/crossentropy": 2.164412271976471,
+      "loss/hidden": 3.22734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17602096050977706,
+      "step": 30600
+    },
+    {
+      "epoch": 0.76525,
+      "grad_norm": 32.75,
+      "grad_norm_var": 8.624739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.301,
+      "loss/crossentropy": 2.0298607796430588,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1785941503942013,
+      "step": 30610
+    },
+    {
+      "epoch": 0.7655,
+      "grad_norm": 34.5,
+      "grad_norm_var": 4.123958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3098,
+      "loss/crossentropy": 2.2510849207639696,
+      "loss/hidden": 3.316796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18415474183857442,
+      "step": 30620
+    },
+    {
+      "epoch": 0.76575,
+      "grad_norm": 29.375,
+      "grad_norm_var": 4.39140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2049,
+      "loss/crossentropy": 2.003419761359692,
+      "loss/hidden": 3.2765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18635775558650494,
+      "step": 30630
+    },
+    {
+      "epoch": 0.766,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.545768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3128,
+      "loss/crossentropy": 2.138288116455078,
+      "loss/hidden": 3.29140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18150971587747336,
+      "step": 30640
+    },
+    {
+      "epoch": 0.76625,
+      "grad_norm": 27.875,
+      "grad_norm_var": 2.600455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2306,
+      "loss/crossentropy": 2.03582369312644,
+      "loss/hidden": 3.309765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19014431536197662,
+      "step": 30650
+    },
+    {
+      "epoch": 0.7665,
+      "grad_norm": 27.5,
+      "grad_norm_var": 5.403059895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.0785,
+      "loss/crossentropy": 2.023620304465294,
+      "loss/hidden": 3.297265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17827179692685605,
+      "step": 30660
+    },
+    {
+      "epoch": 0.76675,
+      "grad_norm": 31.0,
+      "grad_norm_var": 5.3603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2912,
+      "loss/crossentropy": 2.2133360475301744,
+      "loss/hidden": 3.422265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19060195097699761,
+      "step": 30670
+    },
+    {
+      "epoch": 0.767,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.738997395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.1324,
+      "loss/crossentropy": 2.0448534056544303,
+      "loss/hidden": 3.3390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1814136505126953,
+      "step": 30680
+    },
+    {
+      "epoch": 0.76725,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.6879557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2882,
+      "loss/crossentropy": 2.0667665399610997,
+      "loss/hidden": 3.218359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17576975822448732,
+      "step": 30690
+    },
+    {
+      "epoch": 0.7675,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.371809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1785,
+      "loss/crossentropy": 1.9460737973451614,
+      "loss/hidden": 3.28671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16462803408503532,
+      "step": 30700
+    },
+    {
+      "epoch": 0.76775,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.758072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3006,
+      "loss/crossentropy": 2.1625903218984606,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19786368366330861,
+      "step": 30710
+    },
+    {
+      "epoch": 0.768,
+      "grad_norm": 31.625,
+      "grad_norm_var": 4.087955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2984,
+      "loss/crossentropy": 2.115101757645607,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1990343313664198,
+      "step": 30720
+    },
+    {
+      "epoch": 0.76825,
+      "grad_norm": 30.0,
+      "grad_norm_var": 31.29375,
+      "learning_rate": 0.0001,
+      "loss": 7.3308,
+      "loss/crossentropy": 2.1543532446026803,
+      "loss/hidden": 3.354296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18993538860231637,
+      "step": 30730
+    },
+    {
+      "epoch": 0.7685,
+      "grad_norm": 31.625,
+      "grad_norm_var": 13.255208333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3246,
+      "loss/crossentropy": 2.147171225398779,
+      "loss/hidden": 3.284765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18650086652487516,
+      "step": 30740
+    },
+    {
+      "epoch": 0.76875,
+      "grad_norm": 31.75,
+      "grad_norm_var": 13.110872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3292,
+      "loss/crossentropy": 2.050896894186735,
+      "loss/hidden": 3.39140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19013228435069324,
+      "step": 30750
+    },
+    {
+      "epoch": 0.769,
+      "grad_norm": 31.625,
+      "grad_norm_var": 42.296809895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2417,
+      "loss/crossentropy": 1.9304289899766445,
+      "loss/hidden": 3.415625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18646475402638316,
+      "step": 30760
+    },
+    {
+      "epoch": 0.76925,
+      "grad_norm": 31.5,
+      "grad_norm_var": 4.1150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2404,
+      "loss/crossentropy": 2.0934823572635652,
+      "loss/hidden": 3.294140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19288243278861045,
+      "step": 30770
+    },
+    {
+      "epoch": 0.7695,
+      "grad_norm": 28.75,
+      "grad_norm_var": 3.3622395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1472,
+      "loss/crossentropy": 2.119046524167061,
+      "loss/hidden": 3.434765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19379116278141736,
+      "step": 30780
+    },
+    {
+      "epoch": 0.76975,
+      "grad_norm": 28.125,
+      "grad_norm_var": 3.0747395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.228,
+      "loss/crossentropy": 1.8516628809273243,
+      "loss/hidden": 3.355859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17867155242711305,
+      "step": 30790
+    },
+    {
+      "epoch": 0.77,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.3854166666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2606,
+      "loss/crossentropy": 2.140632301568985,
+      "loss/hidden": 3.315625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17682827562093734,
+      "step": 30800
+    },
+    {
+      "epoch": 0.77025,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.33515625,
+      "learning_rate": 0.0001,
+      "loss": 7.1309,
+      "loss/crossentropy": 2.063164585828781,
+      "loss/hidden": 3.331640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18230299968272448,
+      "step": 30810
+    },
+    {
+      "epoch": 0.7705,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.8103515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2604,
+      "loss/crossentropy": 2.110166022181511,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18783345203846694,
+      "step": 30820
+    },
+    {
+      "epoch": 0.77075,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.004166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.298,
+      "loss/crossentropy": 2.1085711009800434,
+      "loss/hidden": 3.364453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2007377685047686,
+      "step": 30830
+    },
+    {
+      "epoch": 0.771,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.7343098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2494,
+      "loss/crossentropy": 1.923082959651947,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17613047789782285,
+      "step": 30840
+    },
+    {
+      "epoch": 0.77125,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.986167998798452e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.245,
+      "loss/crossentropy": 2.139000003039837,
+      "loss/hidden": 3.369140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18553725760430098,
+      "step": 30850
+    },
+    {
+      "epoch": 0.7715,
+      "grad_norm": 31.75,
+      "grad_norm_var": 6.801041666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2407,
+      "loss/crossentropy": 1.9860042557120323,
+      "loss/hidden": 3.293359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16922438573092222,
+      "step": 30860
+    },
+    {
+      "epoch": 0.77175,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.1455729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3205,
+      "loss/crossentropy": 2.2089149236679075,
+      "loss/hidden": 3.340625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19413873814046384,
+      "step": 30870
+    },
+    {
+      "epoch": 0.772,
+      "grad_norm": 26.125,
+      "grad_norm_var": 3.0122395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2108,
+      "loss/crossentropy": 2.0838850632309915,
+      "loss/hidden": 3.22265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1679495433345437,
+      "step": 30880
+    },
+    {
+      "epoch": 0.77225,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.1375,
+      "learning_rate": 0.0001,
+      "loss": 7.1952,
+      "loss/crossentropy": 2.0687290251255037,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17775915022939442,
+      "step": 30890
+    },
+    {
+      "epoch": 0.7725,
+      "grad_norm": 28.0,
+      "grad_norm_var": 4.011393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2563,
+      "loss/crossentropy": 1.9918573498725891,
+      "loss/hidden": 3.315234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17511531990021467,
+      "step": 30900
+    },
+    {
+      "epoch": 0.77275,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.5603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2497,
+      "loss/crossentropy": 1.9922025822103024,
+      "loss/hidden": 3.43203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18127510212361814,
+      "step": 30910
+    },
+    {
+      "epoch": 0.773,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.7614583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2854,
+      "loss/crossentropy": 2.0986859157681463,
+      "loss/hidden": 3.355859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1825839588418603,
+      "step": 30920
+    },
+    {
+      "epoch": 0.77325,
+      "grad_norm": 32.75,
+      "grad_norm_var": 5.251822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2091,
+      "loss/crossentropy": 1.9582366243004798,
+      "loss/hidden": 3.43125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1809095237404108,
+      "step": 30930
+    },
+    {
+      "epoch": 0.7735,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.8372395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2826,
+      "loss/crossentropy": 1.9070976704359055,
+      "loss/hidden": 3.349609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1710755567997694,
+      "step": 30940
+    },
+    {
+      "epoch": 0.77375,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.6020833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2084,
+      "loss/crossentropy": 2.0629549741744997,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18181780502200126,
+      "step": 30950
+    },
+    {
+      "epoch": 0.774,
+      "grad_norm": 33.5,
+      "grad_norm_var": 4.3853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.1774,
+      "loss/crossentropy": 2.1738601446151735,
+      "loss/hidden": 3.35,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2032660385593772,
+      "step": 30960
+    },
+    {
+      "epoch": 0.77425,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.38515625,
+      "learning_rate": 0.0001,
+      "loss": 7.317,
+      "loss/crossentropy": 2.079986987262964,
+      "loss/hidden": 3.43828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19429001584649086,
+      "step": 30970
+    },
+    {
+      "epoch": 0.7745,
+      "grad_norm": 31.125,
+      "grad_norm_var": 4.668684895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.1493,
+      "loss/crossentropy": 2.0519056364893915,
+      "loss/hidden": 3.22265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16815223284065722,
+      "step": 30980
+    },
+    {
+      "epoch": 0.77475,
+      "grad_norm": 28.375,
+      "grad_norm_var": 5.291080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2312,
+      "loss/crossentropy": 2.052493926882744,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18330379351973533,
+      "step": 30990
+    },
+    {
+      "epoch": 0.775,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.12265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2823,
+      "loss/crossentropy": 2.086025407910347,
+      "loss/hidden": 3.294921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18335492983460427,
+      "step": 31000
+    },
+    {
+      "epoch": 0.77525,
+      "grad_norm": 29.0,
+      "grad_norm_var": 5.005989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2367,
+      "loss/crossentropy": 2.048542061448097,
+      "loss/hidden": 3.364453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18725468441843987,
+      "step": 31010
+    },
+    {
+      "epoch": 0.7755,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.0184895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3251,
+      "loss/crossentropy": 2.036131452769041,
+      "loss/hidden": 3.347265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1737860631197691,
+      "step": 31020
+    },
+    {
+      "epoch": 0.77575,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.3999348958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.262,
+      "loss/crossentropy": 2.017168144881725,
+      "loss/hidden": 3.3203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.167899782769382,
+      "step": 31030
+    },
+    {
+      "epoch": 0.776,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.342643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.1701,
+      "loss/crossentropy": 2.019613781571388,
+      "loss/hidden": 3.36953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19059567637741565,
+      "step": 31040
+    },
+    {
+      "epoch": 0.77625,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.4270833333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2972,
+      "loss/crossentropy": 2.0828250303864477,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18261822033673525,
+      "step": 31050
+    },
+    {
+      "epoch": 0.7765,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.8910807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.326,
+      "loss/crossentropy": 2.0209526136517524,
+      "loss/hidden": 3.322265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17059254441410304,
+      "step": 31060
+    },
+    {
+      "epoch": 0.77675,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.356705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2019,
+      "loss/crossentropy": 2.1125204801559447,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17937722243368626,
+      "step": 31070
+    },
+    {
+      "epoch": 0.777,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.4181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.1838,
+      "loss/crossentropy": 2.2028726637363434,
+      "loss/hidden": 3.256640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17989998534321786,
+      "step": 31080
+    },
+    {
+      "epoch": 0.77725,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.671875,
+      "learning_rate": 0.0001,
+      "loss": 7.1709,
+      "loss/crossentropy": 2.1625898659229277,
+      "loss/hidden": 3.390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19090434685349464,
+      "step": 31090
+    },
+    {
+      "epoch": 0.7775,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.36640625,
+      "learning_rate": 0.0001,
+      "loss": 7.1326,
+      "loss/crossentropy": 2.036789299547672,
+      "loss/hidden": 3.3421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17714376896619796,
+      "step": 31100
+    },
+    {
+      "epoch": 0.77775,
+      "grad_norm": 31.75,
+      "grad_norm_var": 11.40390625,
+      "learning_rate": 0.0001,
+      "loss": 7.1735,
+      "loss/crossentropy": 1.9948045410215856,
+      "loss/hidden": 3.24921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1798159122467041,
+      "step": 31110
+    },
+    {
+      "epoch": 0.778,
+      "grad_norm": 34.5,
+      "grad_norm_var": 4.034309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2609,
+      "loss/crossentropy": 2.1702212870121,
+      "loss/hidden": 3.223828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17852376736700534,
+      "step": 31120
+    },
+    {
+      "epoch": 0.77825,
+      "grad_norm": 31.625,
+      "grad_norm_var": 4.676822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3185,
+      "loss/crossentropy": 2.0733341470360758,
+      "loss/hidden": 3.31328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18980380473658442,
+      "step": 31130
+    },
+    {
+      "epoch": 0.7785,
+      "grad_norm": 31.25,
+      "grad_norm_var": 4.430143229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.1541,
+      "loss/crossentropy": 1.9170797616243362,
+      "loss/hidden": 3.327734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1943414082750678,
+      "step": 31140
+    },
+    {
+      "epoch": 0.77875,
+      "grad_norm": 29.125,
+      "grad_norm_var": 14.3228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.0941,
+      "loss/crossentropy": 2.0785103261470796,
+      "loss/hidden": 3.370703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2052849479019642,
+      "step": 31150
+    },
+    {
+      "epoch": 0.779,
+      "grad_norm": 28.625,
+      "grad_norm_var": 4.9869140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3521,
+      "loss/crossentropy": 2.0296490490436554,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1873609971255064,
+      "step": 31160
+    },
+    {
+      "epoch": 0.77925,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.9150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2165,
+      "loss/crossentropy": 2.070235808193684,
+      "loss/hidden": 3.379296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18882376216351987,
+      "step": 31170
+    },
+    {
+      "epoch": 0.7795,
+      "grad_norm": 28.0,
+      "grad_norm_var": 4.44765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2705,
+      "loss/crossentropy": 1.8607801221311093,
+      "loss/hidden": 3.296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16972522204741836,
+      "step": 31180
+    },
+    {
+      "epoch": 0.77975,
+      "grad_norm": 32.5,
+      "grad_norm_var": 5.82265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3687,
+      "loss/crossentropy": 2.151284599304199,
+      "loss/hidden": 3.364453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1864160757511854,
+      "step": 31190
+    },
+    {
+      "epoch": 0.78,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.4530598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2288,
+      "loss/crossentropy": 2.014324263483286,
+      "loss/hidden": 3.339453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18028491344302894,
+      "step": 31200
+    },
+    {
+      "epoch": 0.78025,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.230208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2867,
+      "loss/crossentropy": 2.024941012263298,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1874180819839239,
+      "step": 31210
+    },
+    {
+      "epoch": 0.7805,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.124739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3255,
+      "loss/crossentropy": 2.0621088325977324,
+      "loss/hidden": 3.3125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18965848479419947,
+      "step": 31220
+    },
+    {
+      "epoch": 0.78075,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.1369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.1658,
+      "loss/crossentropy": 2.0059763975441456,
+      "loss/hidden": 3.42578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20813378430902957,
+      "step": 31230
+    },
+    {
+      "epoch": 0.781,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.903125,
+      "learning_rate": 0.0001,
+      "loss": 7.3986,
+      "loss/crossentropy": 2.168729566037655,
+      "loss/hidden": 3.34296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19630307257175444,
+      "step": 31240
+    },
+    {
+      "epoch": 0.78125,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.8056640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3344,
+      "loss/crossentropy": 2.1517349526286127,
+      "loss/hidden": 3.301171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18015162888914346,
+      "step": 31250
+    },
+    {
+      "epoch": 0.7815,
+      "grad_norm": 31.625,
+      "grad_norm_var": 14.411458333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2129,
+      "loss/crossentropy": 2.06531667560339,
+      "loss/hidden": 3.431640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19317530654370785,
+      "step": 31260
+    },
+    {
+      "epoch": 0.78175,
+      "grad_norm": 38.25,
+      "grad_norm_var": 7.57265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3526,
+      "loss/crossentropy": 2.1732255190610887,
+      "loss/hidden": 3.3203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19884103927761315,
+      "step": 31270
+    },
+    {
+      "epoch": 0.782,
+      "grad_norm": 30.5,
+      "grad_norm_var": 8.1947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2598,
+      "loss/crossentropy": 1.9770694464445113,
+      "loss/hidden": 3.264453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18086720816791058,
+      "step": 31280
+    },
+    {
+      "epoch": 0.78225,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.708268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3301,
+      "loss/crossentropy": 2.0416681550443174,
+      "loss/hidden": 3.30390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17893482595682145,
+      "step": 31290
+    },
+    {
+      "epoch": 0.7825,
+      "grad_norm": 32.75,
+      "grad_norm_var": 8.041666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2272,
+      "loss/crossentropy": 1.9394095972180367,
+      "loss/hidden": 3.362890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1878746159374714,
+      "step": 31300
+    },
+    {
+      "epoch": 0.78275,
+      "grad_norm": 36.5,
+      "grad_norm_var": 8.885872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2945,
+      "loss/crossentropy": 2.1827884405851363,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1965128371492028,
+      "step": 31310
+    },
+    {
+      "epoch": 0.783,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.603125,
+      "learning_rate": 0.0001,
+      "loss": 7.2984,
+      "loss/crossentropy": 2.0775409430265426,
+      "loss/hidden": 3.24140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17661921344697476,
+      "step": 31320
+    },
+    {
+      "epoch": 0.78325,
+      "grad_norm": 30.625,
+      "grad_norm_var": 8.294205729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.1817,
+      "loss/crossentropy": 2.049132463335991,
+      "loss/hidden": 3.316796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1787800071761012,
+      "step": 31330
+    },
+    {
+      "epoch": 0.7835,
+      "grad_norm": 32.0,
+      "grad_norm_var": 4.706184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3231,
+      "loss/crossentropy": 2.022278678417206,
+      "loss/hidden": 3.328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19319646768271923,
+      "step": 31340
+    },
+    {
+      "epoch": 0.78375,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.845572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3237,
+      "loss/crossentropy": 1.9787674874067307,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18886952772736548,
+      "step": 31350
+    },
+    {
+      "epoch": 0.784,
+      "grad_norm": 29.5,
+      "grad_norm_var": 4.2744140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3666,
+      "loss/crossentropy": 2.1626318007707597,
+      "loss/hidden": 3.330078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19823330659419297,
+      "step": 31360
+    },
+    {
+      "epoch": 0.78425,
+      "grad_norm": 27.75,
+      "grad_norm_var": 4.5025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2335,
+      "loss/crossentropy": 2.2147083409130572,
+      "loss/hidden": 3.284765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18974209930747749,
+      "step": 31370
+    },
+    {
+      "epoch": 0.7845,
+      "grad_norm": 29.75,
+      "grad_norm_var": 5.703059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2754,
+      "loss/crossentropy": 1.9638444483280182,
+      "loss/hidden": 3.210546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1664100494235754,
+      "step": 31380
+    },
+    {
+      "epoch": 0.78475,
+      "grad_norm": 28.625,
+      "grad_norm_var": 5.27265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2791,
+      "loss/crossentropy": 2.1053778350353243,
+      "loss/hidden": 3.345703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18106228299438953,
+      "step": 31390
+    },
+    {
+      "epoch": 0.785,
+      "grad_norm": 30.25,
+      "grad_norm_var": 56.23515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3181,
+      "loss/crossentropy": 2.079754628241062,
+      "loss/hidden": 3.3890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19201560821384192,
+      "step": 31400
+    },
+    {
+      "epoch": 0.78525,
+      "grad_norm": 30.5,
+      "grad_norm_var": 31.951041666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2104,
+      "loss/crossentropy": 2.1266677528619766,
+      "loss/hidden": 3.327734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18508619628846645,
+      "step": 31410
+    },
+    {
+      "epoch": 0.7855,
+      "grad_norm": 32.0,
+      "grad_norm_var": 35.484309895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.0706,
+      "loss/crossentropy": 1.9031139247119426,
+      "loss/hidden": 3.255078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16362121552228928,
+      "step": 31420
+    },
+    {
+      "epoch": 0.78575,
+      "grad_norm": 31.125,
+      "grad_norm_var": 10.720833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1675,
+      "loss/crossentropy": 2.0395915627479555,
+      "loss/hidden": 3.315625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17914918782189487,
+      "step": 31430
+    },
+    {
+      "epoch": 0.786,
+      "grad_norm": 27.375,
+      "grad_norm_var": 15.347330729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2693,
+      "loss/crossentropy": 2.0384490221738814,
+      "loss/hidden": 3.349609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17985808495432137,
+      "step": 31440
+    },
+    {
+      "epoch": 0.78625,
+      "grad_norm": 33.0,
+      "grad_norm_var": 10.926822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1775,
+      "loss/crossentropy": 2.0808909833431244,
+      "loss/hidden": 3.24609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17647454431280493,
+      "step": 31450
+    },
+    {
+      "epoch": 0.7865,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.29375,
+      "learning_rate": 0.0001,
+      "loss": 7.2395,
+      "loss/crossentropy": 2.2060162991285326,
+      "loss/hidden": 3.312109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18310822043567895,
+      "step": 31460
+    },
+    {
+      "epoch": 0.78675,
+      "grad_norm": 28.75,
+      "grad_norm_var": 372.24140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2611,
+      "loss/crossentropy": 1.9570839531719684,
+      "loss/hidden": 3.308203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1838856378570199,
+      "step": 31470
+    },
+    {
+      "epoch": 0.787,
+      "grad_norm": 28.375,
+      "grad_norm_var": 5.662955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2132,
+      "loss/crossentropy": 1.8940647289156913,
+      "loss/hidden": 3.36328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18987012309953571,
+      "step": 31480
+    },
+    {
+      "epoch": 0.78725,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.489518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2335,
+      "loss/crossentropy": 2.2421848446130754,
+      "loss/hidden": 3.32109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18429996352642775,
+      "step": 31490
+    },
+    {
+      "epoch": 0.7875,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.330989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3104,
+      "loss/crossentropy": 2.0037948578596114,
+      "loss/hidden": 3.36328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18187961634248495,
+      "step": 31500
+    },
+    {
+      "epoch": 0.78775,
+      "grad_norm": 29.375,
+      "grad_norm_var": 7.914583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2202,
+      "loss/crossentropy": 1.9032418720424176,
+      "loss/hidden": 3.430078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1838343466632068,
+      "step": 31510
+    },
+    {
+      "epoch": 0.788,
+      "grad_norm": 27.375,
+      "grad_norm_var": 4.1603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2262,
+      "loss/crossentropy": 2.0772057265043258,
+      "loss/hidden": 3.3734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19711360968649388,
+      "step": 31520
+    },
+    {
+      "epoch": 0.78825,
+      "grad_norm": 35.5,
+      "grad_norm_var": 9.198372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3303,
+      "loss/crossentropy": 2.2158702582120897,
+      "loss/hidden": 3.263671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18437313195317984,
+      "step": 31530
+    },
+    {
+      "epoch": 0.7885,
+      "grad_norm": 35.75,
+      "grad_norm_var": 7.027083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.302,
+      "loss/crossentropy": 2.0729110084474085,
+      "loss/hidden": 3.371484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19737910926342012,
+      "step": 31540
+    },
+    {
+      "epoch": 0.78875,
+      "grad_norm": 32.25,
+      "grad_norm_var": 6.404166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2099,
+      "loss/crossentropy": 1.9750752970576286,
+      "loss/hidden": 3.21640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17290026945993303,
+      "step": 31550
+    },
+    {
+      "epoch": 0.789,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.4629557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2183,
+      "loss/crossentropy": 2.0953447185456753,
+      "loss/hidden": 3.398828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1804077338427305,
+      "step": 31560
+    },
+    {
+      "epoch": 0.78925,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.66015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2443,
+      "loss/crossentropy": 1.9566784024238586,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18815564569085835,
+      "step": 31570
+    },
+    {
+      "epoch": 0.7895,
+      "grad_norm": 33.75,
+      "grad_norm_var": 9.175,
+      "learning_rate": 0.0001,
+      "loss": 7.3692,
+      "loss/crossentropy": 2.2338188737630844,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18873065207153558,
+      "step": 31580
+    },
+    {
+      "epoch": 0.78975,
+      "grad_norm": 29.25,
+      "grad_norm_var": 9.987239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2562,
+      "loss/crossentropy": 2.080914391577244,
+      "loss/hidden": 3.29296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17150079626590015,
+      "step": 31590
+    },
+    {
+      "epoch": 0.79,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.3212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2378,
+      "loss/crossentropy": 2.064169317483902,
+      "loss/hidden": 3.45390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20276217330247165,
+      "step": 31600
+    },
+    {
+      "epoch": 0.79025,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.957291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3602,
+      "loss/crossentropy": 2.1499217487871647,
+      "loss/hidden": 3.398046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18527807723730802,
+      "step": 31610
+    },
+    {
+      "epoch": 0.7905,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.027083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3215,
+      "loss/crossentropy": 2.1228356897830962,
+      "loss/hidden": 3.25703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17830363363027574,
+      "step": 31620
+    },
+    {
+      "epoch": 0.79075,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.4160807291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2688,
+      "loss/crossentropy": 2.0881614208221437,
+      "loss/hidden": 3.258984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17270825430750847,
+      "step": 31630
+    },
+    {
+      "epoch": 0.791,
+      "grad_norm": 28.625,
+      "grad_norm_var": 3.7525390625,
+      "learning_rate": 0.0001,
+      "loss": 7.2279,
+      "loss/crossentropy": 1.9683130264282227,
+      "loss/hidden": 3.426953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17459874097257852,
+      "step": 31640
+    },
+    {
+      "epoch": 0.79125,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.24375,
+      "learning_rate": 0.0001,
+      "loss": 7.2418,
+      "loss/crossentropy": 1.9514197140932084,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18445694111287594,
+      "step": 31650
+    },
+    {
+      "epoch": 0.7915,
+      "grad_norm": 32.5,
+      "grad_norm_var": 4.755143229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.1507,
+      "loss/crossentropy": 2.079793556034565,
+      "loss/hidden": 3.2609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17297195587307215,
+      "step": 31660
+    },
+    {
+      "epoch": 0.79175,
+      "grad_norm": 33.0,
+      "grad_norm_var": 4.960872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4178,
+      "loss/crossentropy": 2.0618962317705156,
+      "loss/hidden": 3.32421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19019667170941829,
+      "step": 31670
+    },
+    {
+      "epoch": 0.792,
+      "grad_norm": 30.625,
+      "grad_norm_var": 5.7509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2239,
+      "loss/crossentropy": 1.9956995896995067,
+      "loss/hidden": 3.40234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17896303534507751,
+      "step": 31680
+    },
+    {
+      "epoch": 0.79225,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.070572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1065,
+      "loss/crossentropy": 2.023432979732752,
+      "loss/hidden": 3.29453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1705683045089245,
+      "step": 31690
+    },
+    {
+      "epoch": 0.7925,
+      "grad_norm": 27.125,
+      "grad_norm_var": 2.5697265625,
+      "learning_rate": 0.0001,
+      "loss": 7.1217,
+      "loss/crossentropy": 1.946426709741354,
+      "loss/hidden": 3.348046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17420749654993414,
+      "step": 31700
+    },
+    {
+      "epoch": 0.79275,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.3197265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2364,
+      "loss/crossentropy": 2.0409046217799185,
+      "loss/hidden": 3.396875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18621115908026695,
+      "step": 31710
+    },
+    {
+      "epoch": 0.793,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.3207682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2649,
+      "loss/crossentropy": 2.114277198910713,
+      "loss/hidden": 3.3,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18227699343115092,
+      "step": 31720
+    },
+    {
+      "epoch": 0.79325,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.3583333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1856,
+      "loss/crossentropy": 1.9716971188783645,
+      "loss/hidden": 3.4265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1855311430990696,
+      "step": 31730
+    },
+    {
+      "epoch": 0.7935,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.7712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2383,
+      "loss/crossentropy": 2.037404325604439,
+      "loss/hidden": 3.34609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1694024408236146,
+      "step": 31740
+    },
+    {
+      "epoch": 0.79375,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.5403116401424246e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2469,
+      "loss/crossentropy": 1.9700333438813686,
+      "loss/hidden": 3.295703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1846833111718297,
+      "step": 31750
+    },
+    {
+      "epoch": 0.794,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.5403116394849674e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.1773,
+      "loss/crossentropy": 2.212307733297348,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18856476843357087,
+      "step": 31760
+    },
+    {
+      "epoch": 0.79425,
+      "grad_norm": 41.25,
+      "grad_norm_var": 14.876822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.256,
+      "loss/crossentropy": 2.0969568237662317,
+      "loss/hidden": 3.275390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.179645280726254,
+      "step": 31770
+    },
+    {
+      "epoch": 0.7945,
+      "grad_norm": 28.875,
+      "grad_norm_var": 10.3712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.1819,
+      "loss/crossentropy": 1.9347251377999783,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18520953366532922,
+      "step": 31780
+    },
+    {
+      "epoch": 0.79475,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.5327473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2494,
+      "loss/crossentropy": 2.0052943006157875,
+      "loss/hidden": 3.30390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17678212691098452,
+      "step": 31790
+    },
+    {
+      "epoch": 0.795,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.6518229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.251,
+      "loss/crossentropy": 1.9200729496777058,
+      "loss/hidden": 3.412890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1845573834143579,
+      "step": 31800
+    },
+    {
+      "epoch": 0.79525,
+      "grad_norm": 29.875,
+      "grad_norm_var": 6.238541666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3009,
+      "loss/crossentropy": 1.982697883248329,
+      "loss/hidden": 3.38671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19470857754349707,
+      "step": 31810
+    },
+    {
+      "epoch": 0.7955,
+      "grad_norm": 30.625,
+      "grad_norm_var": 6.095768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2466,
+      "loss/crossentropy": 1.9890022471547126,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1903789708390832,
+      "step": 31820
+    },
+    {
+      "epoch": 0.79575,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.4806640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2873,
+      "loss/crossentropy": 1.9891909763216973,
+      "loss/hidden": 3.441796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1799360180273652,
+      "step": 31830
+    },
+    {
+      "epoch": 0.796,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.8962890625,
+      "learning_rate": 0.0001,
+      "loss": 7.237,
+      "loss/crossentropy": 2.1192200325429438,
+      "loss/hidden": 3.292578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18091769460588694,
+      "step": 31840
+    },
+    {
+      "epoch": 0.79625,
+      "grad_norm": 28.75,
+      "grad_norm_var": 1.8872395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1984,
+      "loss/crossentropy": 2.197053752094507,
+      "loss/hidden": 3.341015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19096684809774161,
+      "step": 31850
+    },
+    {
+      "epoch": 0.7965,
+      "grad_norm": 34.75,
+      "grad_norm_var": 1.8240281960738632e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3018,
+      "loss/crossentropy": 1.9654372721910476,
+      "loss/hidden": 3.380859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18191756587475538,
+      "step": 31860
+    },
+    {
+      "epoch": 0.79675,
+      "grad_norm": 30.375,
+      "grad_norm_var": 4.390625,
+      "learning_rate": 0.0001,
+      "loss": 7.1343,
+      "loss/crossentropy": 2.1827359586954116,
+      "loss/hidden": 3.388671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19493103586137295,
+      "step": 31870
+    },
+    {
+      "epoch": 0.797,
+      "grad_norm": 35.0,
+      "grad_norm_var": 4.96875,
+      "learning_rate": 0.0001,
+      "loss": 7.2734,
+      "loss/crossentropy": 2.0850317165255547,
+      "loss/hidden": 3.4046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1965682774782181,
+      "step": 31880
+    },
+    {
+      "epoch": 0.79725,
+      "grad_norm": 29.0,
+      "grad_norm_var": 4.255989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2898,
+      "loss/crossentropy": 2.078572702407837,
+      "loss/hidden": 3.38984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19365447759628296,
+      "step": 31890
+    },
+    {
+      "epoch": 0.7975,
+      "grad_norm": 29.875,
+      "grad_norm_var": 5.2650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3819,
+      "loss/crossentropy": 2.0934070348739624,
+      "loss/hidden": 3.403515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17799469958990813,
+      "step": 31900
+    },
+    {
+      "epoch": 0.79775,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.856705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2383,
+      "loss/crossentropy": 1.9600264072418212,
+      "loss/hidden": 3.501171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1934324972331524,
+      "step": 31910
+    },
+    {
+      "epoch": 0.798,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.471809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2629,
+      "loss/crossentropy": 2.2288677148520946,
+      "loss/hidden": 3.417578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20668185902759434,
+      "step": 31920
+    },
+    {
+      "epoch": 0.79825,
+      "grad_norm": 32.75,
+      "grad_norm_var": 6.603125,
+      "learning_rate": 0.0001,
+      "loss": 7.2431,
+      "loss/crossentropy": 2.189570128917694,
+      "loss/hidden": 3.434765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19532648399472236,
+      "step": 31930
+    },
+    {
+      "epoch": 0.7985,
+      "grad_norm": 34.0,
+      "grad_norm_var": 10.4603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.246,
+      "loss/crossentropy": 2.052780894935131,
+      "loss/hidden": 3.2640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18245863653719424,
+      "step": 31940
+    },
+    {
+      "epoch": 0.79875,
+      "grad_norm": 30.25,
+      "grad_norm_var": 11.5791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2474,
+      "loss/crossentropy": 2.1735198058187963,
+      "loss/hidden": 3.386328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18904183022677898,
+      "step": 31950
+    },
+    {
+      "epoch": 0.799,
+      "grad_norm": 29.875,
+      "grad_norm_var": 6.057291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2311,
+      "loss/crossentropy": 1.9964329220354557,
+      "loss/hidden": 3.341796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17117326706647873,
+      "step": 31960
+    },
+    {
+      "epoch": 0.79925,
+      "grad_norm": 28.875,
+      "grad_norm_var": 7.140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2959,
+      "loss/crossentropy": 2.0637815829366444,
+      "loss/hidden": 3.348046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1760326312389225,
+      "step": 31970
+    },
+    {
+      "epoch": 0.7995,
+      "grad_norm": 31.875,
+      "grad_norm_var": 5.26015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2168,
+      "loss/crossentropy": 1.9845145612955093,
+      "loss/hidden": 3.405078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1851344184949994,
+      "step": 31980
+    },
+    {
+      "epoch": 0.79975,
+      "grad_norm": 31.5,
+      "grad_norm_var": 3.655989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2525,
+      "loss/crossentropy": 2.039809539914131,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17651782967150212,
+      "step": 31990
+    },
+    {
+      "epoch": 0.8,
+      "grad_norm": 32.0,
+      "grad_norm_var": 5.589322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2411,
+      "loss/crossentropy": 2.051901635527611,
+      "loss/hidden": 3.36171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17814793512225152,
+      "step": 32000
+    },
+    {
+      "epoch": 0.80025,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.0666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2601,
+      "loss/crossentropy": 2.167706046998501,
+      "loss/hidden": 3.4578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20020014122128488,
+      "step": 32010
+    },
+    {
+      "epoch": 0.8005,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.7509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2548,
+      "loss/crossentropy": 2.031809562444687,
+      "loss/hidden": 3.4171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1915106700733304,
+      "step": 32020
+    },
+    {
+      "epoch": 0.80075,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.5479166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3371,
+      "loss/crossentropy": 2.151665323972702,
+      "loss/hidden": 3.242578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17666382808238268,
+      "step": 32030
+    },
+    {
+      "epoch": 0.801,
+      "grad_norm": 34.0,
+      "grad_norm_var": 3.2744140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2745,
+      "loss/crossentropy": 2.155677234381437,
+      "loss/hidden": 3.317578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1825503976084292,
+      "step": 32040
+    },
+    {
+      "epoch": 0.80125,
+      "grad_norm": 29.375,
+      "grad_norm_var": 4.258333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2659,
+      "loss/crossentropy": 2.0210466548800468,
+      "loss/hidden": 3.380859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18113178480416536,
+      "step": 32050
+    },
+    {
+      "epoch": 0.8015,
+      "grad_norm": 28.75,
+      "grad_norm_var": 4.313997395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.1235,
+      "loss/crossentropy": 1.9424000725150108,
+      "loss/hidden": 3.414453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1696411618962884,
+      "step": 32060
+    },
+    {
+      "epoch": 0.80175,
+      "grad_norm": 30.375,
+      "grad_norm_var": 6.124934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2674,
+      "loss/crossentropy": 2.0733315840363504,
+      "loss/hidden": 3.38046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18977922163903713,
+      "step": 32070
+    },
+    {
+      "epoch": 0.802,
+      "grad_norm": 29.25,
+      "grad_norm_var": 7.609309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.259,
+      "loss/crossentropy": 2.227187469601631,
+      "loss/hidden": 3.248828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17911691386252643,
+      "step": 32080
+    },
+    {
+      "epoch": 0.80225,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.6754557291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.234,
+      "loss/crossentropy": 2.087716729938984,
+      "loss/hidden": 3.321484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18047652654349805,
+      "step": 32090
+    },
+    {
+      "epoch": 0.8025,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.0400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4021,
+      "loss/crossentropy": 2.131810285151005,
+      "loss/hidden": 3.276953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17247401773929597,
+      "step": 32100
+    },
+    {
+      "epoch": 0.80275,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.55390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3817,
+      "loss/crossentropy": 2.1549403719604014,
+      "loss/hidden": 3.381640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18837472070008515,
+      "step": 32110
+    },
+    {
+      "epoch": 0.803,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.9442057291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.1035,
+      "loss/crossentropy": 1.9505949556827544,
+      "loss/hidden": 3.394140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18204545034095646,
+      "step": 32120
+    },
+    {
+      "epoch": 0.80325,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.92890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2371,
+      "loss/crossentropy": 2.1348301008343697,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18901300579309463,
+      "step": 32130
+    },
+    {
+      "epoch": 0.8035,
+      "grad_norm": 33.25,
+      "grad_norm_var": 3.1728515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2579,
+      "loss/crossentropy": 1.9484988793730735,
+      "loss/hidden": 3.271484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16889106966555117,
+      "step": 32140
+    },
+    {
+      "epoch": 0.80375,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.8393229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2884,
+      "loss/crossentropy": 1.9826452344655991,
+      "loss/hidden": 3.376953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1743983631953597,
+      "step": 32150
+    },
+    {
+      "epoch": 0.804,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.5322916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.1942,
+      "loss/crossentropy": 1.9426303870975972,
+      "loss/hidden": 3.42890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1772203067317605,
+      "step": 32160
+    },
+    {
+      "epoch": 0.80425,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.874934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2248,
+      "loss/crossentropy": 1.870173018425703,
+      "loss/hidden": 3.34453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1688270630314946,
+      "step": 32170
+    },
+    {
+      "epoch": 0.8045,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.5759765625,
+      "learning_rate": 0.0001,
+      "loss": 7.1157,
+      "loss/crossentropy": 1.9939643263816833,
+      "loss/hidden": 3.325,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17504618484526874,
+      "step": 32180
+    },
+    {
+      "epoch": 0.80475,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.2587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2281,
+      "loss/crossentropy": 2.026793968677521,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17457820419222117,
+      "step": 32190
+    },
+    {
+      "epoch": 0.805,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.758736220941137e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2813,
+      "loss/crossentropy": 1.8412863582372665,
+      "loss/hidden": 3.398828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18289455082267522,
+      "step": 32200
+    },
+    {
+      "epoch": 0.80525,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.758736220602027e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3068,
+      "loss/crossentropy": 2.1275208979845046,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.182334803044796,
+      "step": 32210
+    },
+    {
+      "epoch": 0.8055,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.758736219923808e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3141,
+      "loss/crossentropy": 2.134224483370781,
+      "loss/hidden": 3.266015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17503854436799884,
+      "step": 32220
+    },
+    {
+      "epoch": 0.80575,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.758736219342478e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4317,
+      "loss/crossentropy": 2.066732983291149,
+      "loss/hidden": 3.4828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20872533712536095,
+      "step": 32230
+    },
+    {
+      "epoch": 0.806,
+      "grad_norm": 26.625,
+      "grad_norm_var": 5.002018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2406,
+      "loss/crossentropy": 1.9790171518921853,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17605258710682392,
+      "step": 32240
+    },
+    {
+      "epoch": 0.80625,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.5551432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3137,
+      "loss/crossentropy": 2.115340569615364,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.191914526745677,
+      "step": 32250
+    },
+    {
+      "epoch": 0.8065,
+      "grad_norm": 33.0,
+      "grad_norm_var": 1.95625,
+      "learning_rate": 0.0001,
+      "loss": 7.2523,
+      "loss/crossentropy": 1.9939750552177429,
+      "loss/hidden": 3.313671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18340199431404472,
+      "step": 32260
+    },
+    {
+      "epoch": 0.80675,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.0518229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2693,
+      "loss/crossentropy": 2.1312461979687214,
+      "loss/hidden": 3.22109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17243078984320165,
+      "step": 32270
+    },
+    {
+      "epoch": 0.807,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.7080729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2228,
+      "loss/crossentropy": 1.9979605212807656,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1727626511827111,
+      "step": 32280
+    },
+    {
+      "epoch": 0.80725,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.4259765625,
+      "learning_rate": 0.0001,
+      "loss": 7.1558,
+      "loss/crossentropy": 2.046182955801487,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1849920243024826,
+      "step": 32290
+    },
+    {
+      "epoch": 0.8075,
+      "grad_norm": 28.25,
+      "grad_norm_var": 3.084830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2157,
+      "loss/crossentropy": 2.107586032152176,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18789408281445502,
+      "step": 32300
+    },
+    {
+      "epoch": 0.80775,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.986458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.135,
+      "loss/crossentropy": 1.966269464045763,
+      "loss/hidden": 3.33984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17599091026932,
+      "step": 32310
+    },
+    {
+      "epoch": 0.808,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.6926432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.258,
+      "loss/crossentropy": 2.03958098590374,
+      "loss/hidden": 3.459765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18970332182943822,
+      "step": 32320
+    },
+    {
+      "epoch": 0.80825,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.4893229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2886,
+      "loss/crossentropy": 1.927356332540512,
+      "loss/hidden": 3.398046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18331324979662894,
+      "step": 32330
+    },
+    {
+      "epoch": 0.8085,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.8212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3217,
+      "loss/crossentropy": 2.1595819175243376,
+      "loss/hidden": 3.243359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1796510087326169,
+      "step": 32340
+    },
+    {
+      "epoch": 0.80875,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.135872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1902,
+      "loss/crossentropy": 2.0535615980625153,
+      "loss/hidden": 3.271875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17316078525036574,
+      "step": 32350
+    },
+    {
+      "epoch": 0.809,
+      "grad_norm": 29.5,
+      "grad_norm_var": 4.070833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1575,
+      "loss/crossentropy": 1.870839986205101,
+      "loss/hidden": 3.29375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1662488218396902,
+      "step": 32360
+    },
+    {
+      "epoch": 0.80925,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.912239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1563,
+      "loss/crossentropy": 2.0238643392920492,
+      "loss/hidden": 3.265234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16979822777211667,
+      "step": 32370
+    },
+    {
+      "epoch": 0.8095,
+      "grad_norm": 34.0,
+      "grad_norm_var": 3.70625,
+      "learning_rate": 0.0001,
+      "loss": 7.2548,
+      "loss/crossentropy": 2.030308040976524,
+      "loss/hidden": 3.32265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17933697663247586,
+      "step": 32380
+    },
+    {
+      "epoch": 0.80975,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.705208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1595,
+      "loss/crossentropy": 2.097300034761429,
+      "loss/hidden": 3.277734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1716611221432686,
+      "step": 32390
+    },
+    {
+      "epoch": 0.81,
+      "grad_norm": 27.75,
+      "grad_norm_var": 8.028580729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2592,
+      "loss/crossentropy": 2.1147085294127463,
+      "loss/hidden": 3.287109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17622052859514953,
+      "step": 32400
+    },
+    {
+      "epoch": 0.81025,
+      "grad_norm": 31.875,
+      "grad_norm_var": 4.1619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.1481,
+      "loss/crossentropy": 2.1348901122808455,
+      "loss/hidden": 3.3109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18319515213370324,
+      "step": 32410
+    },
+    {
+      "epoch": 0.8105,
+      "grad_norm": 33.25,
+      "grad_norm_var": 5.037239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2796,
+      "loss/crossentropy": 2.0976379089057446,
+      "loss/hidden": 3.294921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16924114488065242,
+      "step": 32420
+    },
+    {
+      "epoch": 0.81075,
+      "grad_norm": 32.0,
+      "grad_norm_var": 15.3806640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3044,
+      "loss/crossentropy": 2.1474139034748077,
+      "loss/hidden": 3.408203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1934512373059988,
+      "step": 32430
+    },
+    {
+      "epoch": 0.811,
+      "grad_norm": 34.5,
+      "grad_norm_var": 6.9244140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2681,
+      "loss/crossentropy": 2.1129303738474845,
+      "loss/hidden": 3.326953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19321042932569982,
+      "step": 32440
+    },
+    {
+      "epoch": 0.81125,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.9455729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4055,
+      "loss/crossentropy": 2.2073852330446244,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18046779334545135,
+      "step": 32450
+    },
+    {
+      "epoch": 0.8115,
+      "grad_norm": 28.0,
+      "grad_norm_var": 2.247330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2706,
+      "loss/crossentropy": 2.009214510768652,
+      "loss/hidden": 3.44765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20849293209612368,
+      "step": 32460
+    },
+    {
+      "epoch": 0.81175,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.4058471889324826e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4795,
+      "loss/crossentropy": 2.1204657569527625,
+      "loss/hidden": 3.307421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18255203776061535,
+      "step": 32470
+    },
+    {
+      "epoch": 0.812,
+      "grad_norm": 28.375,
+      "grad_norm_var": 3.1010416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.163,
+      "loss/crossentropy": 1.9438684552907943,
+      "loss/hidden": 3.297265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1655712489038706,
+      "step": 32480
+    },
+    {
+      "epoch": 0.81225,
+      "grad_norm": 27.75,
+      "grad_norm_var": 13.784309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2748,
+      "loss/crossentropy": 1.906498458981514,
+      "loss/hidden": 3.275390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17149967923760415,
+      "step": 32490
+    },
+    {
+      "epoch": 0.8125,
+      "grad_norm": 32.75,
+      "grad_norm_var": 18.0337890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2568,
+      "loss/crossentropy": 2.1418545827269555,
+      "loss/hidden": 3.341796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18672561962157488,
+      "step": 32500
+    },
+    {
+      "epoch": 0.81275,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.83515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2452,
+      "loss/crossentropy": 2.0591210328042506,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1714493166655302,
+      "step": 32510
+    },
+    {
+      "epoch": 0.813,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.671875,
+      "learning_rate": 0.0001,
+      "loss": 7.1253,
+      "loss/crossentropy": 2.0408148244023323,
+      "loss/hidden": 3.3375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18097794353961943,
+      "step": 32520
+    },
+    {
+      "epoch": 0.81325,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.7809895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2145,
+      "loss/crossentropy": 1.9985629439353942,
+      "loss/hidden": 3.3296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19131027571856976,
+      "step": 32530
+    },
+    {
+      "epoch": 0.8135,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.2426432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3301,
+      "loss/crossentropy": 2.1079444982111455,
+      "loss/hidden": 3.371484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18785760272294283,
+      "step": 32540
+    },
+    {
+      "epoch": 0.81375,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.5431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.174,
+      "loss/crossentropy": 1.9769691236317157,
+      "loss/hidden": 3.228125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16153431748971342,
+      "step": 32550
+    },
+    {
+      "epoch": 0.814,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.133072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1737,
+      "loss/crossentropy": 1.9369481757283211,
+      "loss/hidden": 3.3890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18391805402934552,
+      "step": 32560
+    },
+    {
+      "epoch": 0.81425,
+      "grad_norm": 31.375,
+      "grad_norm_var": 5.5322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3592,
+      "loss/crossentropy": 2.074393917620182,
+      "loss/hidden": 3.2703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17730166036635636,
+      "step": 32570
+    },
+    {
+      "epoch": 0.8145,
+      "grad_norm": 30.25,
+      "grad_norm_var": 7.649934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2958,
+      "loss/crossentropy": 2.152645838260651,
+      "loss/hidden": 3.2671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1779523069038987,
+      "step": 32580
+    },
+    {
+      "epoch": 0.81475,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.687239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2586,
+      "loss/crossentropy": 2.1111764326691627,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2009077413007617,
+      "step": 32590
+    },
+    {
+      "epoch": 0.815,
+      "grad_norm": 31.0,
+      "grad_norm_var": 0.853125,
+      "learning_rate": 0.0001,
+      "loss": 7.2938,
+      "loss/crossentropy": 2.095028135180473,
+      "loss/hidden": 3.312109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18357501607388257,
+      "step": 32600
+    },
+    {
+      "epoch": 0.81525,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.2643229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3045,
+      "loss/crossentropy": 2.18403902053833,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20085975714027882,
+      "step": 32610
+    },
+    {
+      "epoch": 0.8155,
+      "grad_norm": 7549747200.0,
+      "grad_norm_var": 3.562417644801884e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3539,
+      "loss/crossentropy": 2.0945954963564875,
+      "loss/hidden": 3.355078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19835255537182092,
+      "step": 32620
+    },
+    {
+      "epoch": 0.81575,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.562417644180603e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2437,
+      "loss/crossentropy": 2.1617246508598327,
+      "loss/hidden": 3.331640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1850573167204857,
+      "step": 32630
+    },
+    {
+      "epoch": 0.816,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.133268229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2513,
+      "loss/crossentropy": 2.0281495213508607,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17307685893028973,
+      "step": 32640
+    },
+    {
+      "epoch": 0.81625,
+      "grad_norm": 31.375,
+      "grad_norm_var": 10.743489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1824,
+      "loss/crossentropy": 1.9506276100873947,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17777946423739194,
+      "step": 32650
+    },
+    {
+      "epoch": 0.8165,
+      "grad_norm": 30.625,
+      "grad_norm_var": 9.271809895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.1952,
+      "loss/crossentropy": 2.1222976088523864,
+      "loss/hidden": 3.339453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18664605617523194,
+      "step": 32660
+    },
+    {
+      "epoch": 0.81675,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.0072265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2998,
+      "loss/crossentropy": 2.107238310575485,
+      "loss/hidden": 3.324609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19150303788483142,
+      "step": 32670
+    },
+    {
+      "epoch": 0.817,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.9150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3482,
+      "loss/crossentropy": 1.946744628250599,
+      "loss/hidden": 3.34140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18452872475609183,
+      "step": 32680
+    },
+    {
+      "epoch": 0.81725,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.1708333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2419,
+      "loss/crossentropy": 2.169080337882042,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18254951611161233,
+      "step": 32690
+    },
+    {
+      "epoch": 0.8175,
+      "grad_norm": 29.0,
+      "grad_norm_var": 2.3052083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2529,
+      "loss/crossentropy": 2.2490514785051348,
+      "loss/hidden": 3.312109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1880893513560295,
+      "step": 32700
+    },
+    {
+      "epoch": 0.81775,
+      "grad_norm": 28.625,
+      "grad_norm_var": 20.976497395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3429,
+      "loss/crossentropy": 2.1124695405364036,
+      "loss/hidden": 3.41171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1869946725666523,
+      "step": 32710
+    },
+    {
+      "epoch": 0.818,
+      "grad_norm": 28.625,
+      "grad_norm_var": 21.052083333333332,
+      "learning_rate": 0.0001,
+      "loss": 7.2308,
+      "loss/crossentropy": 1.9733390748500823,
+      "loss/hidden": 3.284765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17016562707722188,
+      "step": 32720
+    },
+    {
+      "epoch": 0.81825,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.2416666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2682,
+      "loss/crossentropy": 2.0912339106202125,
+      "loss/hidden": 3.35078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18757227193564177,
+      "step": 32730
+    },
+    {
+      "epoch": 0.8185,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.9098307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2908,
+      "loss/crossentropy": 2.046810332685709,
+      "loss/hidden": 3.30703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1753769489005208,
+      "step": 32740
+    },
+    {
+      "epoch": 0.81875,
+      "grad_norm": 28.875,
+      "grad_norm_var": 85.02682291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.1812,
+      "loss/crossentropy": 1.8806068405508996,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1857995780184865,
+      "step": 32750
+    },
+    {
+      "epoch": 0.819,
+      "grad_norm": 31.75,
+      "grad_norm_var": 83.22916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2393,
+      "loss/crossentropy": 2.263414332270622,
+      "loss/hidden": 3.304296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19157245978713036,
+      "step": 32760
+    },
+    {
+      "epoch": 0.81925,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.247330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2049,
+      "loss/crossentropy": 2.0481057576835155,
+      "loss/hidden": 3.237109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16489372327923774,
+      "step": 32770
+    },
+    {
+      "epoch": 0.8195,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.637955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3102,
+      "loss/crossentropy": 2.0778881147503854,
+      "loss/hidden": 3.4078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18638940406963228,
+      "step": 32780
+    },
+    {
+      "epoch": 0.81975,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.450455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2894,
+      "loss/crossentropy": 2.0030703216791155,
+      "loss/hidden": 3.303515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17577996235340834,
+      "step": 32790
+    },
+    {
+      "epoch": 0.82,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.9561848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1411,
+      "loss/crossentropy": 1.983315745741129,
+      "loss/hidden": 3.34453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1809986751526594,
+      "step": 32800
+    },
+    {
+      "epoch": 0.82025,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.5622395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.201,
+      "loss/crossentropy": 2.0666268527507783,
+      "loss/hidden": 3.298046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17382380943745374,
+      "step": 32810
+    },
+    {
+      "epoch": 0.8205,
+      "grad_norm": 28.625,
+      "grad_norm_var": 8.472916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.217,
+      "loss/crossentropy": 1.9913094595074654,
+      "loss/hidden": 3.451171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17583483047783374,
+      "step": 32820
+    },
+    {
+      "epoch": 0.82075,
+      "grad_norm": 29.625,
+      "grad_norm_var": 9.392708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2831,
+      "loss/crossentropy": 2.131434071063995,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18594115655869245,
+      "step": 32830
+    },
+    {
+      "epoch": 0.821,
+      "grad_norm": 28.25,
+      "grad_norm_var": 4.114583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.162,
+      "loss/crossentropy": 2.108327967301011,
+      "loss/hidden": 3.3171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18347832215949894,
+      "step": 32840
+    },
+    {
+      "epoch": 0.82125,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.1426432291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.1914,
+      "loss/crossentropy": 1.989778284728527,
+      "loss/hidden": 3.370703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1808457243256271,
+      "step": 32850
+    },
+    {
+      "epoch": 0.8215,
+      "grad_norm": 28.25,
+      "grad_norm_var": 2.7124348958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2049,
+      "loss/crossentropy": 2.120199552178383,
+      "loss/hidden": 3.325,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18457971159368752,
+      "step": 32860
+    },
+    {
+      "epoch": 0.82175,
+      "grad_norm": 35.5,
+      "grad_norm_var": 25.367708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2266,
+      "loss/crossentropy": 2.0434179216623307,
+      "loss/hidden": 3.268359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1738625967875123,
+      "step": 32870
+    },
+    {
+      "epoch": 0.822,
+      "grad_norm": 28.875,
+      "grad_norm_var": 20.552018229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3397,
+      "loss/crossentropy": 2.1949902296066286,
+      "loss/hidden": 3.3859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20381880030035973,
+      "step": 32880
+    },
+    {
+      "epoch": 0.82225,
+      "grad_norm": 30.625,
+      "grad_norm_var": 20.99375,
+      "learning_rate": 0.0001,
+      "loss": 7.2043,
+      "loss/crossentropy": 2.0969431951642035,
+      "loss/hidden": 3.19375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16738458890467883,
+      "step": 32890
+    },
+    {
+      "epoch": 0.8225,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.6,
+      "learning_rate": 0.0001,
+      "loss": 7.3717,
+      "loss/crossentropy": 1.910037212073803,
+      "loss/hidden": 3.496875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21138516459614037,
+      "step": 32900
+    },
+    {
+      "epoch": 0.82275,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.47890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2944,
+      "loss/crossentropy": 2.128745178878307,
+      "loss/hidden": 3.290234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16938637476414442,
+      "step": 32910
+    },
+    {
+      "epoch": 0.823,
+      "grad_norm": 31.125,
+      "grad_norm_var": 4.466080729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3261,
+      "loss/crossentropy": 2.1102588385343553,
+      "loss/hidden": 3.233984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18096004836261273,
+      "step": 32920
+    },
+    {
+      "epoch": 0.82325,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.6264973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2731,
+      "loss/crossentropy": 2.0781460791826247,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20577163994312286,
+      "step": 32930
+    },
+    {
+      "epoch": 0.8235,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.6613932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3275,
+      "loss/crossentropy": 2.0820566445589064,
+      "loss/hidden": 3.337890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17738431375473737,
+      "step": 32940
+    },
+    {
+      "epoch": 0.82375,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.8696271412957893e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3092,
+      "loss/crossentropy": 2.1891760557889937,
+      "loss/hidden": 3.32265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18127844985574484,
+      "step": 32950
+    },
+    {
+      "epoch": 0.824,
+      "grad_norm": 31.75,
+      "grad_norm_var": 6.083333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.243,
+      "loss/crossentropy": 2.1428692400455476,
+      "loss/hidden": 3.33046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19357558973133565,
+      "step": 32960
+    },
+    {
+      "epoch": 0.82425,
+      "grad_norm": 28.875,
+      "grad_norm_var": 5.588997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2652,
+      "loss/crossentropy": 1.9545904472470284,
+      "loss/hidden": 3.45625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18726088060066104,
+      "step": 32970
+    },
+    {
+      "epoch": 0.8245,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.6593098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2011,
+      "loss/crossentropy": 2.19513940513134,
+      "loss/hidden": 3.315234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18281609732657672,
+      "step": 32980
+    },
+    {
+      "epoch": 0.82475,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.0087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.1861,
+      "loss/crossentropy": 1.9000049561262131,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18603194374591112,
+      "step": 32990
+    },
+    {
+      "epoch": 0.825,
+      "grad_norm": 34.5,
+      "grad_norm_var": 5.230143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3087,
+      "loss/crossentropy": 2.176528325676918,
+      "loss/hidden": 3.380859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19508822299540043,
+      "step": 33000
+    },
+    {
+      "epoch": 0.82525,
+      "grad_norm": 29.75,
+      "grad_norm_var": 6.870768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.294,
+      "loss/crossentropy": 2.074225351214409,
+      "loss/hidden": 3.287890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1698618305847049,
+      "step": 33010
+    },
+    {
+      "epoch": 0.8255,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.5259765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2318,
+      "loss/crossentropy": 1.9874062195420266,
+      "loss/hidden": 3.309375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17083260375075043,
+      "step": 33020
+    },
+    {
+      "epoch": 0.82575,
+      "grad_norm": 34.0,
+      "grad_norm_var": 17.623958333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2262,
+      "loss/crossentropy": 2.009798040986061,
+      "loss/hidden": 3.38359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17537497971206903,
+      "step": 33030
+    },
+    {
+      "epoch": 0.826,
+      "grad_norm": 31.875,
+      "grad_norm_var": 4.7509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.247,
+      "loss/crossentropy": 1.975242407619953,
+      "loss/hidden": 3.297265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1771410683169961,
+      "step": 33040
+    },
+    {
+      "epoch": 0.82625,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.6885416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3327,
+      "loss/crossentropy": 2.1277087360620497,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18781726695597173,
+      "step": 33050
+    },
+    {
+      "epoch": 0.8265,
+      "grad_norm": 29.5,
+      "grad_norm_var": 19.176041666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2487,
+      "loss/crossentropy": 2.0511118680238725,
+      "loss/hidden": 3.383203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18265996258705855,
+      "step": 33060
+    },
+    {
+      "epoch": 0.82675,
+      "grad_norm": 28.875,
+      "grad_norm_var": 21.9181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2396,
+      "loss/crossentropy": 2.0069188207387922,
+      "loss/hidden": 3.42109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18895241040736438,
+      "step": 33070
+    },
+    {
+      "epoch": 0.827,
+      "grad_norm": 29.0,
+      "grad_norm_var": 5.089518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.1739,
+      "loss/crossentropy": 2.0877312146127225,
+      "loss/hidden": 3.300390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1880142292007804,
+      "step": 33080
+    },
+    {
+      "epoch": 0.82725,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.7997395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2304,
+      "loss/crossentropy": 2.1121929422020913,
+      "loss/hidden": 3.38203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18759419713169337,
+      "step": 33090
+    },
+    {
+      "epoch": 0.8275,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.0629557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2302,
+      "loss/crossentropy": 2.0372960224747656,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1759640196338296,
+      "step": 33100
+    },
+    {
+      "epoch": 0.82775,
+      "grad_norm": 28.625,
+      "grad_norm_var": 1.12265625,
+      "learning_rate": 0.0001,
+      "loss": 7.1867,
+      "loss/crossentropy": 2.0432544320821764,
+      "loss/hidden": 3.247265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17888994030654431,
+      "step": 33110
+    },
+    {
+      "epoch": 0.828,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.155042767301509e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3383,
+      "loss/crossentropy": 1.9840383768081664,
+      "loss/hidden": 3.459375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1890584809705615,
+      "step": 33120
+    },
+    {
+      "epoch": 0.82825,
+      "grad_norm": 29.625,
+      "grad_norm_var": 1.7580729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1764,
+      "loss/crossentropy": 1.9866980195045472,
+      "loss/hidden": 3.280859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16517556998878719,
+      "step": 33130
+    },
+    {
+      "epoch": 0.8285,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.1416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2066,
+      "loss/crossentropy": 2.0350388549268246,
+      "loss/hidden": 3.298046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17439607214182615,
+      "step": 33140
+    },
+    {
+      "epoch": 0.82875,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.4358723958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2689,
+      "loss/crossentropy": 2.1262506932020186,
+      "loss/hidden": 3.244921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17747630644589663,
+      "step": 33150
+    },
+    {
+      "epoch": 0.829,
+      "grad_norm": 31.0,
+      "grad_norm_var": 9.006184895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2443,
+      "loss/crossentropy": 2.1790440320968627,
+      "loss/hidden": 3.275390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18086491283029318,
+      "step": 33160
+    },
+    {
+      "epoch": 0.82925,
+      "grad_norm": 28.75,
+      "grad_norm_var": 10.318684895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3319,
+      "loss/crossentropy": 2.174597004055977,
+      "loss/hidden": 3.269140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1854943221434951,
+      "step": 33170
+    },
+    {
+      "epoch": 0.8295,
+      "grad_norm": 34.5,
+      "grad_norm_var": 2.9228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2453,
+      "loss/crossentropy": 2.1312096804380416,
+      "loss/hidden": 3.279296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17487460002303123,
+      "step": 33180
+    },
+    {
+      "epoch": 0.82975,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.530989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2681,
+      "loss/crossentropy": 2.1367484860122206,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18738685622811319,
+      "step": 33190
+    },
+    {
+      "epoch": 0.83,
+      "grad_norm": 28.5,
+      "grad_norm_var": 1.828125,
+      "learning_rate": 0.0001,
+      "loss": 7.228,
+      "loss/crossentropy": 2.1254208497703075,
+      "loss/hidden": 3.281640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1795659614726901,
+      "step": 33200
+    },
+    {
+      "epoch": 0.83025,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.9530598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2011,
+      "loss/crossentropy": 1.9877605877816678,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18452598052099348,
+      "step": 33210
+    },
+    {
+      "epoch": 0.8305,
+      "grad_norm": 35.25,
+      "grad_norm_var": 3.9431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.1594,
+      "loss/crossentropy": 2.0039492681622506,
+      "loss/hidden": 3.334375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18381198737770318,
+      "step": 33220
+    },
+    {
+      "epoch": 0.83075,
+      "grad_norm": 43.5,
+      "grad_norm_var": 14.514322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2984,
+      "loss/crossentropy": 2.1737947434186937,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17826213724911213,
+      "step": 33230
+    },
+    {
+      "epoch": 0.831,
+      "grad_norm": 28.125,
+      "grad_norm_var": 13.200455729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.1889,
+      "loss/crossentropy": 2.1893670335412025,
+      "loss/hidden": 3.312890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1756696652621031,
+      "step": 33240
+    },
+    {
+      "epoch": 0.83125,
+      "grad_norm": 32.25,
+      "grad_norm_var": 6.585416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2065,
+      "loss/crossentropy": 2.2689021706581114,
+      "loss/hidden": 3.353515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18847335111349822,
+      "step": 33250
+    },
+    {
+      "epoch": 0.8315,
+      "grad_norm": 29.875,
+      "grad_norm_var": 4.733072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.273,
+      "loss/crossentropy": 2.1217512801289558,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18651887271553277,
+      "step": 33260
+    },
+    {
+      "epoch": 0.83175,
+      "grad_norm": 32.75,
+      "grad_norm_var": 4.162434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2661,
+      "loss/crossentropy": 1.952786525338888,
+      "loss/hidden": 3.212109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.15934878159314395,
+      "step": 33270
+    },
+    {
+      "epoch": 0.832,
+      "grad_norm": 28.875,
+      "grad_norm_var": 18.381184895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.1678,
+      "loss/crossentropy": 2.1539489194750785,
+      "loss/hidden": 3.331640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1839644392952323,
+      "step": 33280
+    },
+    {
+      "epoch": 0.83225,
+      "grad_norm": 31.5,
+      "grad_norm_var": 6.798958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.295,
+      "loss/crossentropy": 2.144645670056343,
+      "loss/hidden": 3.29765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17679207548499107,
+      "step": 33290
+    },
+    {
+      "epoch": 0.8325,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.0150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3755,
+      "loss/crossentropy": 2.0700169153511525,
+      "loss/hidden": 3.3234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18612122694030403,
+      "step": 33300
+    },
+    {
+      "epoch": 0.83275,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.7768229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1297,
+      "loss/crossentropy": 2.0893400743603707,
+      "loss/hidden": 3.269140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17862566062249244,
+      "step": 33310
+    },
+    {
+      "epoch": 0.833,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.01015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2992,
+      "loss/crossentropy": 2.1281204342842104,
+      "loss/hidden": 3.38125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19392694123089313,
+      "step": 33320
+    },
+    {
+      "epoch": 0.83325,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.66640625,
+      "learning_rate": 0.0001,
+      "loss": 7.266,
+      "loss/crossentropy": 1.9223125174641609,
+      "loss/hidden": 3.365625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1853881362825632,
+      "step": 33330
+    },
+    {
+      "epoch": 0.8335,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.8166015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2632,
+      "loss/crossentropy": 1.9951855912804604,
+      "loss/hidden": 3.263671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17544143386185168,
+      "step": 33340
+    },
+    {
+      "epoch": 0.83375,
+      "grad_norm": 27.5,
+      "grad_norm_var": 1.9083333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2816,
+      "loss/crossentropy": 1.8624406516551972,
+      "loss/hidden": 3.326953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.166407653875649,
+      "step": 33350
+    },
+    {
+      "epoch": 0.834,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.018684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3705,
+      "loss/crossentropy": 2.019158259034157,
+      "loss/hidden": 3.377734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1928328016772866,
+      "step": 33360
+    },
+    {
+      "epoch": 0.83425,
+      "grad_norm": 31.625,
+      "grad_norm_var": 4.66015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2398,
+      "loss/crossentropy": 2.002238343656063,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18211230151355268,
+      "step": 33370
+    },
+    {
+      "epoch": 0.8345,
+      "grad_norm": 29.75,
+      "grad_norm_var": 5.110872395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2584,
+      "loss/crossentropy": 2.167521375417709,
+      "loss/hidden": 3.292578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17853855043649675,
+      "step": 33380
+    },
+    {
+      "epoch": 0.83475,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.7447916666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.1261,
+      "loss/crossentropy": 2.034111428260803,
+      "loss/hidden": 3.320703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18338973429054023,
+      "step": 33390
+    },
+    {
+      "epoch": 0.835,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.1705729166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2059,
+      "loss/crossentropy": 1.937609364837408,
+      "loss/hidden": 3.3984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18285921812057496,
+      "step": 33400
+    },
+    {
+      "epoch": 0.83525,
+      "grad_norm": 28.625,
+      "grad_norm_var": 2.6988932291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.222,
+      "loss/crossentropy": 2.170501431822777,
+      "loss/hidden": 3.253515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17850385289639234,
+      "step": 33410
+    },
+    {
+      "epoch": 0.8355,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.5369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.2907,
+      "loss/crossentropy": 1.869483571499586,
+      "loss/hidden": 3.462109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19331245701760053,
+      "step": 33420
+    },
+    {
+      "epoch": 0.83575,
+      "grad_norm": 33.5,
+      "grad_norm_var": 3.5559895833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2744,
+      "loss/crossentropy": 2.044697883725166,
+      "loss/hidden": 3.322265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17986318413168192,
+      "step": 33430
+    },
+    {
+      "epoch": 0.836,
+      "grad_norm": 31.125,
+      "grad_norm_var": 5.074739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1971,
+      "loss/crossentropy": 2.074133487045765,
+      "loss/hidden": 3.2578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17584548909217118,
+      "step": 33440
+    },
+    {
+      "epoch": 0.83625,
+      "grad_norm": 32.5,
+      "grad_norm_var": 4.56875,
+      "learning_rate": 0.0001,
+      "loss": 7.2264,
+      "loss/crossentropy": 2.1718404933810236,
+      "loss/hidden": 3.39296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1941602316685021,
+      "step": 33450
+    },
+    {
+      "epoch": 0.8365,
+      "grad_norm": 28.5,
+      "grad_norm_var": 3.4833333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3675,
+      "loss/crossentropy": 2.105057214200497,
+      "loss/hidden": 3.2921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1713842570781708,
+      "step": 33460
+    },
+    {
+      "epoch": 0.83675,
+      "grad_norm": 28.0,
+      "grad_norm_var": 2.684375,
+      "learning_rate": 0.0001,
+      "loss": 7.176,
+      "loss/crossentropy": 2.1097921282052994,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20385158769786357,
+      "step": 33470
+    },
+    {
+      "epoch": 0.837,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.5916666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.1873,
+      "loss/crossentropy": 2.0274341255426407,
+      "loss/hidden": 3.309375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17628777120262384,
+      "step": 33480
+    },
+    {
+      "epoch": 0.83725,
+      "grad_norm": 33.75,
+      "grad_norm_var": 3.06015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2902,
+      "loss/crossentropy": 2.1492093369364738,
+      "loss/hidden": 3.3078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18647899404168128,
+      "step": 33490
+    },
+    {
+      "epoch": 0.8375,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.1518229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2008,
+      "loss/crossentropy": 2.2044859319925307,
+      "loss/hidden": 3.3125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1958764422684908,
+      "step": 33500
+    },
+    {
+      "epoch": 0.83775,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.1791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2588,
+      "loss/crossentropy": 1.9938185468316079,
+      "loss/hidden": 3.40625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17682940270751715,
+      "step": 33510
+    },
+    {
+      "epoch": 0.838,
+      "grad_norm": 28.625,
+      "grad_norm_var": 3.0916015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2187,
+      "loss/crossentropy": 1.8948181390762329,
+      "loss/hidden": 3.33046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17874590419232844,
+      "step": 33520
+    },
+    {
+      "epoch": 0.83825,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.660416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2563,
+      "loss/crossentropy": 1.9899358443915844,
+      "loss/hidden": 3.34453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17318380083888768,
+      "step": 33530
+    },
+    {
+      "epoch": 0.8385,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.2639973958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.1997,
+      "loss/crossentropy": 2.09279451072216,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19197166245430708,
+      "step": 33540
+    },
+    {
+      "epoch": 0.83875,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.8114583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2935,
+      "loss/crossentropy": 2.2584090396761893,
+      "loss/hidden": 3.321875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20011573657393456,
+      "step": 33550
+    },
+    {
+      "epoch": 0.839,
+      "grad_norm": 29.0,
+      "grad_norm_var": 1.4705729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2413,
+      "loss/crossentropy": 2.064199483394623,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18448180723935365,
+      "step": 33560
+    },
+    {
+      "epoch": 0.83925,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.6811848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2521,
+      "loss/crossentropy": 2.1389085978269575,
+      "loss/hidden": 3.287109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17851551342755556,
+      "step": 33570
+    },
+    {
+      "epoch": 0.8395,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.9624348958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2863,
+      "loss/crossentropy": 2.1424932688474656,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18908166233450174,
+      "step": 33580
+    },
+    {
+      "epoch": 0.83975,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.928465630524028e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2873,
+      "loss/crossentropy": 1.968450340628624,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18586775716394185,
+      "step": 33590
+    },
+    {
+      "epoch": 0.84,
+      "grad_norm": 29.0,
+      "grad_norm_var": 25.168489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2015,
+      "loss/crossentropy": 2.125929144024849,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1928325628861785,
+      "step": 33600
+    },
+    {
+      "epoch": 0.84025,
+      "grad_norm": 33.0,
+      "grad_norm_var": 30.908268229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3396,
+      "loss/crossentropy": 2.133786876499653,
+      "loss/hidden": 3.336328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18853625562041998,
+      "step": 33610
+    },
+    {
+      "epoch": 0.8405,
+      "grad_norm": 31.25,
+      "grad_norm_var": 8.537434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2523,
+      "loss/crossentropy": 2.0643911197781564,
+      "loss/hidden": 3.2890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16706958692520857,
+      "step": 33620
+    },
+    {
+      "epoch": 0.84075,
+      "grad_norm": 29.125,
+      "grad_norm_var": 15.953580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.253,
+      "loss/crossentropy": 2.180924139916897,
+      "loss/hidden": 3.282421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1812012242153287,
+      "step": 33630
+    },
+    {
+      "epoch": 0.841,
+      "grad_norm": 27.75,
+      "grad_norm_var": 13.53125,
+      "learning_rate": 0.0001,
+      "loss": 7.3464,
+      "loss/crossentropy": 2.1258861504495146,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2176899950951338,
+      "step": 33640
+    },
+    {
+      "epoch": 0.84125,
+      "grad_norm": 33.75,
+      "grad_norm_var": 9.324739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2401,
+      "loss/crossentropy": 2.0648413375020027,
+      "loss/hidden": 3.336328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1812899900600314,
+      "step": 33650
+    },
+    {
+      "epoch": 0.8415,
+      "grad_norm": 36.0,
+      "grad_norm_var": 6.6509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2464,
+      "loss/crossentropy": 2.2324777007102967,
+      "loss/hidden": 3.27109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18248971588909627,
+      "step": 33660
+    },
+    {
+      "epoch": 0.84175,
+      "grad_norm": 36.25,
+      "grad_norm_var": 6.999934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2403,
+      "loss/crossentropy": 2.233609616756439,
+      "loss/hidden": 3.2484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17655839212238789,
+      "step": 33670
+    },
+    {
+      "epoch": 0.842,
+      "grad_norm": 30.5,
+      "grad_norm_var": 4.526822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.291,
+      "loss/crossentropy": 2.103627168387175,
+      "loss/hidden": 3.370703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20420456351712346,
+      "step": 33680
+    },
+    {
+      "epoch": 0.84225,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.9134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.218,
+      "loss/crossentropy": 2.1048862997442486,
+      "loss/hidden": 3.317578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1857083396986127,
+      "step": 33690
+    },
+    {
+      "epoch": 0.8425,
+      "grad_norm": 30.5,
+      "grad_norm_var": 6.632291666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.241,
+      "loss/crossentropy": 2.1183654844760893,
+      "loss/hidden": 3.26171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17055343594402075,
+      "step": 33700
+    },
+    {
+      "epoch": 0.84275,
+      "grad_norm": 30.375,
+      "grad_norm_var": 45.45618489583333,
+      "learning_rate": 0.0001,
+      "loss": 7.1277,
+      "loss/crossentropy": 1.9154896408319473,
+      "loss/hidden": 3.343359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17995868530124426,
+      "step": 33710
+    },
+    {
+      "epoch": 0.843,
+      "grad_norm": 33.0,
+      "grad_norm_var": 42.3447265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2229,
+      "loss/crossentropy": 1.8783961586654185,
+      "loss/hidden": 3.281640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.15951730422675608,
+      "step": 33720
+    },
+    {
+      "epoch": 0.84325,
+      "grad_norm": 27.875,
+      "grad_norm_var": 4.575,
+      "learning_rate": 0.0001,
+      "loss": 7.1576,
+      "loss/crossentropy": 2.2371614634990693,
+      "loss/hidden": 3.29765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20098869111388923,
+      "step": 33730
+    },
+    {
+      "epoch": 0.8435,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.2259765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2066,
+      "loss/crossentropy": 2.0927552759647368,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17918878868222238,
+      "step": 33740
+    },
+    {
+      "epoch": 0.84375,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.2660807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.186,
+      "loss/crossentropy": 2.0027615509927275,
+      "loss/hidden": 3.446484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19067690251395106,
+      "step": 33750
+    },
+    {
+      "epoch": 0.844,
+      "grad_norm": 26.125,
+      "grad_norm_var": 5.451822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1761,
+      "loss/crossentropy": 2.010700835287571,
+      "loss/hidden": 3.372265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1867529757320881,
+      "step": 33760
+    },
+    {
+      "epoch": 0.84425,
+      "grad_norm": 36.25,
+      "grad_norm_var": 8.837434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2334,
+      "loss/crossentropy": 1.998683289438486,
+      "loss/hidden": 3.231640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17415560763329269,
+      "step": 33770
+    },
+    {
+      "epoch": 0.8445,
+      "grad_norm": 28.5,
+      "grad_norm_var": 7.392122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.1901,
+      "loss/crossentropy": 2.180643618106842,
+      "loss/hidden": 3.4046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18043227680027485,
+      "step": 33780
+    },
+    {
+      "epoch": 0.84475,
+      "grad_norm": 35.0,
+      "grad_norm_var": 43.71432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1083,
+      "loss/crossentropy": 2.000649718940258,
+      "loss/hidden": 3.34140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17601131778210402,
+      "step": 33790
+    },
+    {
+      "epoch": 0.845,
+      "grad_norm": 29.25,
+      "grad_norm_var": 59.41041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.158,
+      "loss/crossentropy": 1.9586811505258084,
+      "loss/hidden": 3.2734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17341443207114934,
+      "step": 33800
+    },
+    {
+      "epoch": 0.84525,
+      "grad_norm": 28.875,
+      "grad_norm_var": 50.3228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.255,
+      "loss/crossentropy": 2.2014808893203734,
+      "loss/hidden": 3.407421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1981699451804161,
+      "step": 33810
+    },
+    {
+      "epoch": 0.8455,
+      "grad_norm": 30.625,
+      "grad_norm_var": 37.398372395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3061,
+      "loss/crossentropy": 2.1012670308351518,
+      "loss/hidden": 3.294140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1775698646903038,
+      "step": 33820
+    },
+    {
+      "epoch": 0.84575,
+      "grad_norm": 28.25,
+      "grad_norm_var": 22.502083333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.1653,
+      "loss/crossentropy": 2.0572401136159897,
+      "loss/hidden": 3.328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.182903066650033,
+      "step": 33830
+    },
+    {
+      "epoch": 0.846,
+      "grad_norm": 42.0,
+      "grad_norm_var": 27.852018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.208,
+      "loss/crossentropy": 1.9303223744034768,
+      "loss/hidden": 3.258984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16786477230489255,
+      "step": 33840
+    },
+    {
+      "epoch": 0.84625,
+      "grad_norm": 30.125,
+      "grad_norm_var": 15.620572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1522,
+      "loss/crossentropy": 2.2037465661764144,
+      "loss/hidden": 3.27421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18479030914604663,
+      "step": 33850
+    },
+    {
+      "epoch": 0.8465,
+      "grad_norm": 30.0,
+      "grad_norm_var": 14.82890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2389,
+      "loss/crossentropy": 2.118402448296547,
+      "loss/hidden": 3.391015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19767413921654226,
+      "step": 33860
+    },
+    {
+      "epoch": 0.84675,
+      "grad_norm": 28.875,
+      "grad_norm_var": 15.980989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2069,
+      "loss/crossentropy": 2.0344385020434856,
+      "loss/hidden": 3.406640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18299854248762132,
+      "step": 33870
+    },
+    {
+      "epoch": 0.847,
+      "grad_norm": 30.0,
+      "grad_norm_var": 13.670572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.1785,
+      "loss/crossentropy": 2.100412330776453,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18360245916992426,
+      "step": 33880
+    },
+    {
+      "epoch": 0.84725,
+      "grad_norm": 28.875,
+      "grad_norm_var": 8.151497395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1457,
+      "loss/crossentropy": 2.0797202557325365,
+      "loss/hidden": 3.353515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1801828995347023,
+      "step": 33890
+    },
+    {
+      "epoch": 0.8475,
+      "grad_norm": 33.25,
+      "grad_norm_var": 7.064583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2686,
+      "loss/crossentropy": 2.143502997606993,
+      "loss/hidden": 3.317578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.191785538662225,
+      "step": 33900
+    },
+    {
+      "epoch": 0.84775,
+      "grad_norm": 32.5,
+      "grad_norm_var": 10.1009765625,
+      "learning_rate": 0.0001,
+      "loss": 7.1032,
+      "loss/crossentropy": 1.9439681842923164,
+      "loss/hidden": 3.23984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.15850237123668193,
+      "step": 33910
+    },
+    {
+      "epoch": 0.848,
+      "grad_norm": 29.75,
+      "grad_norm_var": 8.961458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1842,
+      "loss/crossentropy": 2.141938117146492,
+      "loss/hidden": 3.24921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18331899046897887,
+      "step": 33920
+    },
+    {
+      "epoch": 0.84825,
+      "grad_norm": 29.25,
+      "grad_norm_var": 14.096875,
+      "learning_rate": 0.0001,
+      "loss": 7.2517,
+      "loss/crossentropy": 2.0895468056201936,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18215047754347324,
+      "step": 33930
+    },
+    {
+      "epoch": 0.8485,
+      "grad_norm": 30.0,
+      "grad_norm_var": 7.476822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2402,
+      "loss/crossentropy": 2.0649702072143556,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.193724649772048,
+      "step": 33940
+    },
+    {
+      "epoch": 0.84875,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.956184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1436,
+      "loss/crossentropy": 2.058744602650404,
+      "loss/hidden": 3.469921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1868588777258992,
+      "step": 33950
+    },
+    {
+      "epoch": 0.849,
+      "grad_norm": 33.25,
+      "grad_norm_var": 3.496809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4438,
+      "loss/crossentropy": 2.0169904246926307,
+      "loss/hidden": 3.359765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18656898457556964,
+      "step": 33960
+    },
+    {
+      "epoch": 0.84925,
+      "grad_norm": 43.75,
+      "grad_norm_var": 13.911458333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2415,
+      "loss/crossentropy": 1.9969091698527337,
+      "loss/hidden": 3.411328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19089818820357324,
+      "step": 33970
+    },
+    {
+      "epoch": 0.8495,
+      "grad_norm": 31.375,
+      "grad_norm_var": 16.1212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2844,
+      "loss/crossentropy": 2.1538162916898727,
+      "loss/hidden": 3.29453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18127647042274475,
+      "step": 33980
+    },
+    {
+      "epoch": 0.84975,
+      "grad_norm": 28.125,
+      "grad_norm_var": 2.0686848958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2479,
+      "loss/crossentropy": 1.9829765126109122,
+      "loss/hidden": 3.341796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18695083800703288,
+      "step": 33990
+    },
+    {
+      "epoch": 0.85,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.88515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2118,
+      "loss/crossentropy": 2.14188566505909,
+      "loss/hidden": 3.419140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19902665503323078,
+      "step": 34000
+    },
+    {
+      "epoch": 0.85025,
+      "grad_norm": 82.5,
+      "grad_norm_var": 341.77076822916666,
+      "learning_rate": 0.0001,
+      "loss": 7.3306,
+      "loss/crossentropy": 2.0443466253578664,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18300221683457493,
+      "step": 34010
+    },
+    {
+      "epoch": 0.8505,
+      "grad_norm": 30.25,
+      "grad_norm_var": 394.2884765625,
+      "learning_rate": 0.0001,
+      "loss": 7.1513,
+      "loss/crossentropy": 2.0158767446875574,
+      "loss/hidden": 3.244140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16400964334607124,
+      "step": 34020
+    },
+    {
+      "epoch": 0.85075,
+      "grad_norm": 33.0,
+      "grad_norm_var": 6.6931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2646,
+      "loss/crossentropy": 2.0715938180685045,
+      "loss/hidden": 3.332421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17798414621502162,
+      "step": 34030
+    },
+    {
+      "epoch": 0.851,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.2879557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3028,
+      "loss/crossentropy": 2.0998909667134287,
+      "loss/hidden": 3.227734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17252598106861114,
+      "step": 34040
+    },
+    {
+      "epoch": 0.85125,
+      "grad_norm": 31.375,
+      "grad_norm_var": 0.8093098958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3888,
+      "loss/crossentropy": 2.110021045804024,
+      "loss/hidden": 3.35,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18911605402827264,
+      "step": 34050
+    },
+    {
+      "epoch": 0.8515,
+      "grad_norm": 29.75,
+      "grad_norm_var": 5.635416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.187,
+      "loss/crossentropy": 1.9248976431787015,
+      "loss/hidden": 3.353125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17050397992134095,
+      "step": 34060
+    },
+    {
+      "epoch": 0.85175,
+      "grad_norm": 29.375,
+      "grad_norm_var": 0.9635416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3192,
+      "loss/crossentropy": 2.181374564766884,
+      "loss/hidden": 3.415625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19645040035247802,
+      "step": 34070
+    },
+    {
+      "epoch": 0.852,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.314322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.2283,
+      "loss/crossentropy": 2.0438218235969545,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19108605310320853,
+      "step": 34080
+    },
+    {
+      "epoch": 0.85225,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.176497395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2638,
+      "loss/crossentropy": 2.1723656713962556,
+      "loss/hidden": 3.336328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18390613235533237,
+      "step": 34090
+    },
+    {
+      "epoch": 0.8525,
+      "grad_norm": 28.5,
+      "grad_norm_var": 1.8622395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1693,
+      "loss/crossentropy": 2.0392055988311766,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20077015552669764,
+      "step": 34100
+    },
+    {
+      "epoch": 0.85275,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.6499348958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.268,
+      "loss/crossentropy": 2.0958330132067204,
+      "loss/hidden": 3.334375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1809829636476934,
+      "step": 34110
+    },
+    {
+      "epoch": 0.853,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.8759765625,
+      "learning_rate": 0.0001,
+      "loss": 7.1929,
+      "loss/crossentropy": 2.13548312112689,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18752679266035557,
+      "step": 34120
+    },
+    {
+      "epoch": 0.85325,
+      "grad_norm": 42.25,
+      "grad_norm_var": 10.880989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2939,
+      "loss/crossentropy": 2.2205469399690627,
+      "loss/hidden": 3.340234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20043759122490884,
+      "step": 34130
+    },
+    {
+      "epoch": 0.8535,
+      "grad_norm": 32.0,
+      "grad_norm_var": 10.0806640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3005,
+      "loss/crossentropy": 2.0488852843642236,
+      "loss/hidden": 3.276953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17713683266192676,
+      "step": 34140
+    },
+    {
+      "epoch": 0.85375,
+      "grad_norm": 27.25,
+      "grad_norm_var": 2.3520833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1977,
+      "loss/crossentropy": 2.097453436255455,
+      "loss/hidden": 3.27890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18755655977874994,
+      "step": 34150
+    },
+    {
+      "epoch": 0.854,
+      "grad_norm": 43.5,
+      "grad_norm_var": 12.536393229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2194,
+      "loss/crossentropy": 2.0824636235833167,
+      "loss/hidden": 3.2734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18008290715515612,
+      "step": 34160
+    },
+    {
+      "epoch": 0.85425,
+      "grad_norm": 28.625,
+      "grad_norm_var": 14.074934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2031,
+      "loss/crossentropy": 2.1574205800890924,
+      "loss/hidden": 3.3859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19417616873979568,
+      "step": 34170
+    },
+    {
+      "epoch": 0.8545,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.220833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.2651,
+      "loss/crossentropy": 2.0641315899789334,
+      "loss/hidden": 3.403515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18580467496067285,
+      "step": 34180
+    },
+    {
+      "epoch": 0.85475,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.120247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2528,
+      "loss/crossentropy": 2.1908591121435164,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19210774190723895,
+      "step": 34190
+    },
+    {
+      "epoch": 0.855,
+      "grad_norm": 31.0,
+      "grad_norm_var": 8.9,
+      "learning_rate": 0.0001,
+      "loss": 7.234,
+      "loss/crossentropy": 1.9531598590314387,
+      "loss/hidden": 3.328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16695764372125269,
+      "step": 34200
+    },
+    {
+      "epoch": 0.85525,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.5622564690419953e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2399,
+      "loss/crossentropy": 2.1334476828575135,
+      "loss/hidden": 3.380859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18677760362625123,
+      "step": 34210
+    },
+    {
+      "epoch": 0.8555,
+      "grad_norm": 28.25,
+      "grad_norm_var": 1.5622564690367872e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2243,
+      "loss/crossentropy": 2.1223787307739257,
+      "loss/hidden": 3.287109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.177429954521358,
+      "step": 34220
+    },
+    {
+      "epoch": 0.85575,
+      "grad_norm": 32.25,
+      "grad_norm_var": 6.009375,
+      "learning_rate": 0.0001,
+      "loss": 7.1667,
+      "loss/crossentropy": 2.053752092272043,
+      "loss/hidden": 3.391015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17705130875110625,
+      "step": 34230
+    },
+    {
+      "epoch": 0.856,
+      "grad_norm": 31.25,
+      "grad_norm_var": 9.1212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.2172,
+      "loss/crossentropy": 2.0997056752443313,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1921197187155485,
+      "step": 34240
+    },
+    {
+      "epoch": 0.85625,
+      "grad_norm": 29.0,
+      "grad_norm_var": 14.873893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.356,
+      "loss/crossentropy": 2.250038433074951,
+      "loss/hidden": 3.38515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2058642553165555,
+      "step": 34250
+    },
+    {
+      "epoch": 0.8565,
+      "grad_norm": 34.5,
+      "grad_norm_var": 6.51015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2588,
+      "loss/crossentropy": 2.076680043339729,
+      "loss/hidden": 3.445703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19434863217175008,
+      "step": 34260
+    },
+    {
+      "epoch": 0.85675,
+      "grad_norm": 31.5,
+      "grad_norm_var": 5.5416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2713,
+      "loss/crossentropy": 2.0715027406811712,
+      "loss/hidden": 3.28671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18325763065367937,
+      "step": 34270
+    },
+    {
+      "epoch": 0.857,
+      "grad_norm": 33.5,
+      "grad_norm_var": 6.362955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.2176,
+      "loss/crossentropy": 2.1784741804003716,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19372119139879942,
+      "step": 34280
+    },
+    {
+      "epoch": 0.85725,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.1797421956559012e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.2276,
+      "loss/crossentropy": 1.9267006784677505,
+      "loss/hidden": 3.621484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17005468346178532,
+      "step": 34290
+    },
+    {
+      "epoch": 0.8575,
+      "grad_norm": 36.25,
+      "grad_norm_var": 19.323372395833335,
+      "learning_rate": 0.0001,
+      "loss": 7.2047,
+      "loss/crossentropy": 2.0357692562043668,
+      "loss/hidden": 3.265234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16413889769464732,
+      "step": 34300
+    },
+    {
+      "epoch": 0.85775,
+      "grad_norm": 29.75,
+      "grad_norm_var": 18.8625,
+      "learning_rate": 0.0001,
+      "loss": 7.3689,
+      "loss/crossentropy": 2.137700055539608,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21653785947710275,
+      "step": 34310
+    },
+    {
+      "epoch": 0.858,
+      "grad_norm": 30.0,
+      "grad_norm_var": 6.951822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3168,
+      "loss/crossentropy": 2.1582180485129356,
+      "loss/hidden": 3.3953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18372731544077398,
+      "step": 34320
+    },
+    {
+      "epoch": 0.85825,
+      "grad_norm": 32.5,
+      "grad_norm_var": 5.91015625,
+      "learning_rate": 0.0001,
+      "loss": 7.1411,
+      "loss/crossentropy": 1.911540611833334,
+      "loss/hidden": 3.4,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17052888721227646,
+      "step": 34330
+    },
+    {
+      "epoch": 0.8585,
+      "grad_norm": 33.5,
+      "grad_norm_var": 6.578580729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3572,
+      "loss/crossentropy": 2.364978903532028,
+      "loss/hidden": 3.28671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18875811491161584,
+      "step": 34340
+    },
+    {
+      "epoch": 0.85875,
+      "grad_norm": 28.375,
+      "grad_norm_var": 3.8583333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2133,
+      "loss/crossentropy": 2.141178289055824,
+      "loss/hidden": 3.38515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19198058787733316,
+      "step": 34350
+    },
+    {
+      "epoch": 0.859,
+      "grad_norm": 34.5,
+      "grad_norm_var": 5.176822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3441,
+      "loss/crossentropy": 2.094452814757824,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19298046212643385,
+      "step": 34360
+    },
+    {
+      "epoch": 0.85925,
+      "grad_norm": 29.5,
+      "grad_norm_var": 6.773372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.2284,
+      "loss/crossentropy": 2.1614732801914216,
+      "loss/hidden": 3.36484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18951235935091973,
+      "step": 34370
+    },
+    {
+      "epoch": 0.8595,
+      "grad_norm": 28.125,
+      "grad_norm_var": 8.332291666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2134,
+      "loss/crossentropy": 1.923812200129032,
+      "loss/hidden": 3.33046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17469180338084697,
+      "step": 34380
+    },
+    {
+      "epoch": 0.85975,
+      "grad_norm": 29.125,
+      "grad_norm_var": 6.509309895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2371,
+      "loss/crossentropy": 2.095948604494333,
+      "loss/hidden": 3.309375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18810884784907103,
+      "step": 34390
+    },
+    {
+      "epoch": 0.86,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.5561848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3054,
+      "loss/crossentropy": 2.100370090454817,
+      "loss/hidden": 3.296484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1727097921539098,
+      "step": 34400
+    },
+    {
+      "epoch": 0.86025,
+      "grad_norm": 32.25,
+      "grad_norm_var": 6.81640625,
+      "learning_rate": 0.0001,
+      "loss": 7.2286,
+      "loss/crossentropy": 2.0329787097871304,
+      "loss/hidden": 3.31953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19481676891446115,
+      "step": 34410
+    },
+    {
+      "epoch": 0.8605,
+      "grad_norm": 27.625,
+      "grad_norm_var": 3.1384765625,
+      "learning_rate": 0.0001,
+      "loss": 7.1993,
+      "loss/crossentropy": 2.0509812578558924,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1901794293895364,
+      "step": 34420
+    },
+    {
+      "epoch": 0.86075,
+      "grad_norm": 33.25,
+      "grad_norm_var": 4.1134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2347,
+      "loss/crossentropy": 2.080067491531372,
+      "loss/hidden": 3.213671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16772212591022254,
+      "step": 34430
+    },
+    {
+      "epoch": 0.861,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.3749348958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3447,
+      "loss/crossentropy": 2.1038970395922663,
+      "loss/hidden": 3.358203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19416949711740017,
+      "step": 34440
+    },
+    {
+      "epoch": 0.86125,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.803125,
+      "learning_rate": 0.0001,
+      "loss": 7.1554,
+      "loss/crossentropy": 2.1600521214306356,
+      "loss/hidden": 3.258984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1786617579869926,
+      "step": 34450
+    },
+    {
+      "epoch": 0.8615,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.93125,
+      "learning_rate": 0.0001,
+      "loss": 7.2464,
+      "loss/crossentropy": 2.0537144929170608,
+      "loss/hidden": 3.386328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18485849518328906,
+      "step": 34460
+    },
+    {
+      "epoch": 0.86175,
+      "grad_norm": 28.75,
+      "grad_norm_var": 3.091080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4084,
+      "loss/crossentropy": 2.187089277803898,
+      "loss/hidden": 3.3140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19332499224692584,
+      "step": 34470
+    },
+    {
+      "epoch": 0.862,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.1259765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3543,
+      "loss/crossentropy": 2.1333917580544948,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2026039410382509,
+      "step": 34480
+    },
+    {
+      "epoch": 0.86225,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.695572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3226,
+      "loss/crossentropy": 2.2536570131778717,
+      "loss/hidden": 3.3484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20088705383241176,
+      "step": 34490
+    },
+    {
+      "epoch": 0.8625,
+      "grad_norm": 30.5,
+      "grad_norm_var": 4.008072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.19,
+      "loss/crossentropy": 2.0807801954448224,
+      "loss/hidden": 3.343359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17837361767888069,
+      "step": 34500
+    },
+    {
+      "epoch": 0.86275,
+      "grad_norm": 31.5,
+      "grad_norm_var": 28.599739583333335,
+      "learning_rate": 0.0001,
+      "loss": 7.2724,
+      "loss/crossentropy": 2.045263282954693,
+      "loss/hidden": 3.206640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1748804647475481,
+      "step": 34510
+    },
+    {
+      "epoch": 0.863,
+      "grad_norm": 30.25,
+      "grad_norm_var": 31.26015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2109,
+      "loss/crossentropy": 1.9880013443529605,
+      "loss/hidden": 3.409765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18841168489307164,
+      "step": 34520
+    },
+    {
+      "epoch": 0.86325,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.3353515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2164,
+      "loss/crossentropy": 1.9977822721004486,
+      "loss/hidden": 3.296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1732394440099597,
+      "step": 34530
+    },
+    {
+      "epoch": 0.8635,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.8489583333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3079,
+      "loss/crossentropy": 2.0394112788140775,
+      "loss/hidden": 3.306640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18080272432416677,
+      "step": 34540
+    },
+    {
+      "epoch": 0.86375,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.9479166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3328,
+      "loss/crossentropy": 2.027179962396622,
+      "loss/hidden": 3.326171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17251962665468454,
+      "step": 34550
+    },
+    {
+      "epoch": 0.864,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.6488932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1298,
+      "loss/crossentropy": 2.034211057424545,
+      "loss/hidden": 3.369140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19268830977380275,
+      "step": 34560
+    },
+    {
+      "epoch": 0.86425,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.4447916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3167,
+      "loss/crossentropy": 2.0971597135066986,
+      "loss/hidden": 3.357421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18605808783322572,
+      "step": 34570
+    },
+    {
+      "epoch": 0.8645,
+      "grad_norm": 48.75,
+      "grad_norm_var": 35.84348958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.142,
+      "loss/crossentropy": 2.0621576346457005,
+      "loss/hidden": 3.260546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1724056525155902,
+      "step": 34580
+    },
+    {
+      "epoch": 0.86475,
+      "grad_norm": 28.5,
+      "grad_norm_var": 36.61764322916667,
+      "learning_rate": 0.0001,
+      "loss": 7.2735,
+      "loss/crossentropy": 2.1108843713998793,
+      "loss/hidden": 3.34140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18370629232376814,
+      "step": 34590
+    },
+    {
+      "epoch": 0.865,
+      "grad_norm": 29.875,
+      "grad_norm_var": 4.808072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3068,
+      "loss/crossentropy": 2.1198356598615646,
+      "loss/hidden": 3.2515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17393486481159925,
+      "step": 34600
+    },
+    {
+      "epoch": 0.86525,
+      "grad_norm": 29.25,
+      "grad_norm_var": 7.821875,
+      "learning_rate": 0.0001,
+      "loss": 7.1895,
+      "loss/crossentropy": 2.1428691267967226,
+      "loss/hidden": 3.312890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17750228513032199,
+      "step": 34610
+    },
+    {
+      "epoch": 0.8655,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.534309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.1452,
+      "loss/crossentropy": 2.1150192469358444,
+      "loss/hidden": 3.30390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18156156884506344,
+      "step": 34620
+    },
+    {
+      "epoch": 0.86575,
+      "grad_norm": 47.25,
+      "grad_norm_var": 20.116666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1916,
+      "loss/crossentropy": 2.048045912384987,
+      "loss/hidden": 3.2953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17622407525777817,
+      "step": 34630
+    },
+    {
+      "epoch": 0.866,
+      "grad_norm": 44.5,
+      "grad_norm_var": 32.889322916666664,
+      "learning_rate": 0.0001,
+      "loss": 7.2463,
+      "loss/crossentropy": 2.1702506244182587,
+      "loss/hidden": 3.320703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18493551313877105,
+      "step": 34640
+    },
+    {
+      "epoch": 0.86625,
+      "grad_norm": 29.25,
+      "grad_norm_var": 17.79765625,
+      "learning_rate": 0.0001,
+      "loss": 7.1766,
+      "loss/crossentropy": 2.0337833762168884,
+      "loss/hidden": 3.2890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16715510403737427,
+      "step": 34650
+    },
+    {
+      "epoch": 0.8665,
+      "grad_norm": 32.5,
+      "grad_norm_var": 21.022916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3194,
+      "loss/crossentropy": 2.120563616603613,
+      "loss/hidden": 3.244140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18254042975604534,
+      "step": 34660
+    },
+    {
+      "epoch": 0.86675,
+      "grad_norm": 30.5,
+      "grad_norm_var": 21.049934895833335,
+      "learning_rate": 0.0001,
+      "loss": 7.2838,
+      "loss/crossentropy": 2.179687091708183,
+      "loss/hidden": 3.31171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18392705954611302,
+      "step": 34670
+    },
+    {
+      "epoch": 0.867,
+      "grad_norm": 28.75,
+      "grad_norm_var": 22.86640625,
+      "learning_rate": 0.0001,
+      "loss": 7.1964,
+      "loss/crossentropy": 2.0371907725930214,
+      "loss/hidden": 3.272265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16766862897202373,
+      "step": 34680
+    },
+    {
+      "epoch": 0.86725,
+      "grad_norm": 37.5,
+      "grad_norm_var": 9.8853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.2503,
+      "loss/crossentropy": 2.0847177296876906,
+      "loss/hidden": 3.445703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19687753692269325,
+      "step": 34690
+    },
+    {
+      "epoch": 0.8675,
+      "grad_norm": 30.625,
+      "grad_norm_var": 8.3125,
+      "learning_rate": 0.0001,
+      "loss": 7.2033,
+      "loss/crossentropy": 2.1133716337382795,
+      "loss/hidden": 3.2671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17771089039742946,
+      "step": 34700
+    },
+    {
+      "epoch": 0.86775,
+      "grad_norm": 32.0,
+      "grad_norm_var": 15.16015625,
+      "learning_rate": 0.0001,
+      "loss": 7.1315,
+      "loss/crossentropy": 1.9481457851827144,
+      "loss/hidden": 3.29296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16198133826255798,
+      "step": 34710
+    },
+    {
+      "epoch": 0.868,
+      "grad_norm": 28.625,
+      "grad_norm_var": 9.898893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.217,
+      "loss/crossentropy": 2.067484679818153,
+      "loss/hidden": 3.358984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1796406304463744,
+      "step": 34720
+    },
+    {
+      "epoch": 0.86825,
+      "grad_norm": 32.0,
+      "grad_norm_var": 7.970247395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.2962,
+      "loss/crossentropy": 2.044479449093342,
+      "loss/hidden": 3.288671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1899338317103684,
+      "step": 34730
+    },
+    {
+      "epoch": 0.8685,
+      "grad_norm": 29.625,
+      "grad_norm_var": 10.008072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2071,
+      "loss/crossentropy": 1.9862901911139488,
+      "loss/hidden": 3.302734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16481328159570693,
+      "step": 34740
+    },
+    {
+      "epoch": 0.86875,
+      "grad_norm": 29.625,
+      "grad_norm_var": 6.892122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.1911,
+      "loss/crossentropy": 2.061764293164015,
+      "loss/hidden": 3.1796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17607883377932013,
+      "step": 34750
+    },
+    {
+      "epoch": 0.869,
+      "grad_norm": 31.75,
+      "grad_norm_var": 6.23515625,
+      "learning_rate": 0.0001,
+      "loss": 7.1601,
+      "loss/crossentropy": 1.9567598104476929,
+      "loss/hidden": 3.412890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1676634754985571,
+      "step": 34760
+    },
+    {
+      "epoch": 0.86925,
+      "grad_norm": 32.0,
+      "grad_norm_var": 5.8712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.1628,
+      "loss/crossentropy": 2.0253935903310776,
+      "loss/hidden": 3.38203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2007642241194844,
+      "step": 34770
+    },
+    {
+      "epoch": 0.8695,
+      "grad_norm": 32.0,
+      "grad_norm_var": 5.207291666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.2495,
+      "loss/crossentropy": 2.0621690064668656,
+      "loss/hidden": 3.403515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18934654965996742,
+      "step": 34780
+    },
+    {
+      "epoch": 0.86975,
+      "grad_norm": 29.75,
+      "grad_norm_var": 30.786458333333332,
+      "learning_rate": 0.0001,
+      "loss": 7.3168,
+      "loss/crossentropy": 2.1084874011576176,
+      "loss/hidden": 3.30546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17805544696748257,
+      "step": 34790
+    },
+    {
+      "epoch": 0.87,
+      "grad_norm": 28.25,
+      "grad_norm_var": 38.59765625,
+      "learning_rate": 0.0001,
+      "loss": 7.2599,
+      "loss/crossentropy": 2.0882460571825505,
+      "loss/hidden": 3.2828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.175720988959074,
+      "step": 34800
+    },
+    {
+      "epoch": 0.87025,
+      "grad_norm": 31.375,
+      "grad_norm_var": 19.3400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.1688,
+      "loss/crossentropy": 2.16546451151371,
+      "loss/hidden": 3.298828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17413612268865108,
+      "step": 34810
+    },
+    {
+      "epoch": 0.8705,
+      "grad_norm": 31.625,
+      "grad_norm_var": 12.055208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.271,
+      "loss/crossentropy": 2.0857179775834083,
+      "loss/hidden": 3.468359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1955510875210166,
+      "step": 34820
+    },
+    {
+      "epoch": 0.87075,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.5348307291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.2399,
+      "loss/crossentropy": 2.0899414606392384,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1947469917126,
+      "step": 34830
+    },
+    {
+      "epoch": 0.871,
+      "grad_norm": 31.125,
+      "grad_norm_var": 4.118489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3026,
+      "loss/crossentropy": 1.9977812230587007,
+      "loss/hidden": 3.422265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1887421939522028,
+      "step": 34840
+    },
+    {
+      "epoch": 0.87125,
+      "grad_norm": 32.75,
+      "grad_norm_var": 106.38515625,
+      "learning_rate": 0.0001,
+      "loss": 7.1963,
+      "loss/crossentropy": 2.035346057265997,
+      "loss/hidden": 3.316796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18354044388979673,
+      "step": 34850
+    },
+    {
+      "epoch": 0.8715,
+      "grad_norm": 29.25,
+      "grad_norm_var": 40.138997395833336,
+      "learning_rate": 0.0001,
+      "loss": 7.2506,
+      "loss/crossentropy": 2.055702011287212,
+      "loss/hidden": 3.34296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18968002796173095,
+      "step": 34860
+    },
+    {
+      "epoch": 0.87175,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.1434895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1322,
+      "loss/crossentropy": 2.205520695447922,
+      "loss/hidden": 3.261328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17120692394673825,
+      "step": 34870
+    },
+    {
+      "epoch": 0.872,
+      "grad_norm": 32.0,
+      "grad_norm_var": 13.1791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.2685,
+      "loss/crossentropy": 2.1152933806180956,
+      "loss/hidden": 3.4109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20207914523780346,
+      "step": 34880
+    },
+    {
+      "epoch": 0.87225,
+      "grad_norm": 29.75,
+      "grad_norm_var": 9.956705729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3239,
+      "loss/crossentropy": 2.1282338641583918,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18587686102837325,
+      "step": 34890
+    },
+    {
+      "epoch": 0.8725,
+      "grad_norm": 52.25,
+      "grad_norm_var": 31.803580729166665,
+      "learning_rate": 0.0001,
+      "loss": 7.2091,
+      "loss/crossentropy": 2.1109050989151,
+      "loss/hidden": 3.400390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1893902899697423,
+      "step": 34900
+    },
+    {
+      "epoch": 0.87275,
+      "grad_norm": 52.75,
+      "grad_norm_var": 61.08958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.147,
+      "loss/crossentropy": 2.071830262243748,
+      "loss/hidden": 3.13828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16298614926636218,
+      "step": 34910
+    },
+    {
+      "epoch": 0.873,
+      "grad_norm": 53.25,
+      "grad_norm_var": 67.65598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2023,
+      "loss/crossentropy": 2.1021742850542067,
+      "loss/hidden": 3.459375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20394247360527515,
+      "step": 34920
+    },
+    {
+      "epoch": 0.87325,
+      "grad_norm": 31.375,
+      "grad_norm_var": 36.49108072916667,
+      "learning_rate": 0.0001,
+      "loss": 7.3976,
+      "loss/crossentropy": 2.1242296025156975,
+      "loss/hidden": 3.248046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17494969312101602,
+      "step": 34930
+    },
+    {
+      "epoch": 0.8735,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.8572265625,
+      "learning_rate": 0.0001,
+      "loss": 7.2589,
+      "loss/crossentropy": 1.9055627778172493,
+      "loss/hidden": 3.298828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17225345242768525,
+      "step": 34940
+    },
+    {
+      "epoch": 0.87375,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.5186848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.1511,
+      "loss/crossentropy": 1.9887107498943806,
+      "loss/hidden": 3.29140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17664323821663858,
+      "step": 34950
+    },
+    {
+      "epoch": 0.874,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.4518229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.1873,
+      "loss/crossentropy": 2.1399480305612086,
+      "loss/hidden": 3.304296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17877526078373193,
+      "step": 34960
+    },
+    {
+      "epoch": 0.87425,
+      "grad_norm": 27.875,
+      "grad_norm_var": 4.1478515625,
+      "learning_rate": 0.0001,
+      "loss": 7.1986,
+      "loss/crossentropy": 2.2206870928406715,
+      "loss/hidden": 3.332421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18415752444416283,
+      "step": 34970
+    },
+    {
+      "epoch": 0.8745,
+      "grad_norm": 34.25,
+      "grad_norm_var": 6.489518229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2735,
+      "loss/crossentropy": 2.185173198580742,
+      "loss/hidden": 3.2921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18359190542250872,
+      "step": 34980
+    },
+    {
+      "epoch": 0.87475,
+      "grad_norm": 29.875,
+      "grad_norm_var": 4.262239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3373,
+      "loss/crossentropy": 2.264425238966942,
+      "loss/hidden": 3.309375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18859156258404255,
+      "step": 34990
+    },
+    {
+      "epoch": 0.875,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.75625,
+      "learning_rate": 0.0001,
+      "loss": 7.1984,
+      "loss/crossentropy": 2.123842165619135,
+      "loss/hidden": 3.312109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1855444625020027,
+      "step": 35000
+    }
+  ],
+  "logging_steps": 10,
+  "max_steps": 40000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 9223372036854775807,
+  "save_steps": 5000,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1.0001285112030822e+20,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}