Update layers.8/cfg.json
Browse files- layers.8/cfg.json +26 -26
layers.8/cfg.json
CHANGED
|
@@ -1,28 +1,28 @@
|
|
| 1 |
{
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"normalize_decoder": true,
|
| 3 |
+
"num_latents": 0,
|
| 4 |
+
"k": 64,
|
| 5 |
+
"multi_topk": false,
|
| 6 |
+
"layer": 8,
|
| 7 |
+
"device": "cuda",
|
| 8 |
+
"activation_fn_str": "relu",
|
| 9 |
+
"d_sae": 24576,
|
| 10 |
+
"d_in": 1536,
|
| 11 |
+
"architecture": "topk",
|
| 12 |
+
"apply_b_dec_to_input": false,
|
| 13 |
+
"finetuning_scaling_factor": false,
|
| 14 |
+
"context_size": 1024,
|
| 15 |
+
"model_name": "Qwen/Qwen2.5-1.5B-Instruct",
|
| 16 |
+
"hook_name": "blocks.8.hook_sae_acts_topk",
|
| 17 |
+
"hook_layer": 8,
|
| 18 |
+
"hook_head_index": null,
|
| 19 |
+
"prepend_bos": true,
|
| 20 |
+
"dataset_path": "",
|
| 21 |
+
"dataset_trust_remote_code": false,
|
| 22 |
+
"normalize_activations": "none",
|
| 23 |
+
"dtype": "float32",
|
| 24 |
+
"sae_lens_training_version": "eleuther",
|
| 25 |
+
"activation_fn_kwargs": {},
|
| 26 |
+
"neuronpedia_id": null,
|
| 27 |
+
"model_from_pretrained_kwargs": {}
|
| 28 |
}
|