Improve model metadata

#2
by tomaarsen HF staff - opened
Files changed (3) hide show
  1. README.md +125 -124
  2. config.json +2 -2
  3. configuration_nvembed.py +0 -2
README.md CHANGED
@@ -1,6 +1,7 @@
1
  ---
2
  tags:
3
  - mteb
 
4
  - sentence-transformers
5
  model-index:
6
  - name: NV-Embed-v2
@@ -90,17 +91,17 @@ model-index:
90
  - type: map_at_5
91
  value: 61.027
92
  - type: mrr_at_1
93
- value: 0.0
94
  - type: mrr_at_10
95
- value: 0.0
96
  - type: mrr_at_100
97
- value: 0.0
98
  - type: mrr_at_1000
99
- value: 0.0
100
  - type: mrr_at_3
101
- value: 0.0
102
  - type: mrr_at_5
103
- value: 0.0
104
  - type: ndcg_at_1
105
  value: 46.515
106
  - type: ndcg_at_10
@@ -278,17 +279,17 @@ model-index:
278
  - type: map_at_5
279
  value: 42.317083333333336
280
  - type: mrr_at_1
281
- value: 0.0
282
  - type: mrr_at_10
283
- value: 0.0
284
  - type: mrr_at_100
285
- value: 0.0
286
  - type: mrr_at_1000
287
- value: 0.0
288
  - type: mrr_at_3
289
- value: 0.0
290
  - type: mrr_at_5
291
- value: 0.0
292
  - type: ndcg_at_1
293
  value: 38.30616666666667
294
  - type: ndcg_at_10
@@ -349,17 +350,17 @@ model-index:
349
  - type: map_at_5
350
  value: 31.955
351
  - type: mrr_at_1
352
- value: 0.0
353
  - type: mrr_at_10
354
- value: 0.0
355
  - type: mrr_at_100
356
- value: 0.0
357
  - type: mrr_at_1000
358
- value: 0.0
359
  - type: mrr_at_3
360
- value: 0.0
361
  - type: mrr_at_5
362
- value: 0.0
363
  - type: ndcg_at_1
364
  value: 44.104
365
  - type: ndcg_at_10
@@ -420,19 +421,19 @@ model-index:
420
  - type: map_at_5
421
  value: 21.062
422
  - type: mrr_at_1
423
- value: 0.0
424
  - type: mrr_at_10
425
- value: 0.0
426
  - type: mrr_at_100
427
- value: 0.0
428
  - type: mrr_at_1000
429
- value: 0.0
430
  - type: mrr_at_3
431
- value: 0.0
432
  - type: mrr_at_5
433
- value: 0.0
434
  - type: ndcg_at_1
435
- value: 66.0
436
  - type: ndcg_at_10
437
  value: 53.496
438
  - type: ndcg_at_100
@@ -452,7 +453,7 @@ model-index:
452
  - type: precision_at_1000
453
  value: 2.5940000000000003
454
  - type: precision_at_3
455
- value: 61.0
456
  - type: precision_at_5
457
  value: 54.65
458
  - type: recall_at_1
@@ -510,17 +511,17 @@ model-index:
510
  - type: map_at_5
511
  value: 91.262
512
  - type: mrr_at_1
513
- value: 0.0
514
  - type: mrr_at_10
515
- value: 0.0
516
  - type: mrr_at_100
517
- value: 0.0
518
  - type: mrr_at_1000
519
- value: 0.0
520
  - type: mrr_at_3
521
- value: 0.0
522
  - type: mrr_at_5
523
- value: 0.0
524
  - type: ndcg_at_1
525
  value: 91.20899999999999
526
  - type: ndcg_at_10
@@ -581,17 +582,17 @@ model-index:
581
  - type: map_at_5
582
  value: 55.054
583
  - type: mrr_at_1
584
- value: 0.0
585
  - type: mrr_at_10
586
- value: 0.0
587
  - type: mrr_at_100
588
- value: 0.0
589
  - type: mrr_at_1000
590
- value: 0.0
591
  - type: mrr_at_3
592
- value: 0.0
593
  - type: mrr_at_5
594
- value: 0.0
595
  - type: ndcg_at_1
596
  value: 64.815
597
  - type: ndcg_at_10
@@ -652,17 +653,17 @@ model-index:
652
  - type: map_at_5
653
  value: 78.935
654
  - type: mrr_at_1
655
- value: 0.0
656
  - type: mrr_at_10
657
- value: 0.0
658
  - type: mrr_at_100
659
- value: 0.0
660
  - type: mrr_at_1000
661
- value: 0.0
662
  - type: mrr_at_3
663
- value: 0.0
664
  - type: mrr_at_5
665
- value: 0.0
666
  - type: ndcg_at_1
667
  value: 89.305
668
  - type: ndcg_at_10
@@ -734,65 +735,65 @@ model-index:
734
  type: mteb/msmarco
735
  metrics:
736
  - type: map_at_1
737
- value: 0.0
738
  - type: map_at_10
739
  value: 38.342
740
  - type: map_at_100
741
- value: 0.0
742
  - type: map_at_1000
743
- value: 0.0
744
  - type: map_at_3
745
- value: 0.0
746
  - type: map_at_5
747
- value: 0.0
748
  - type: mrr_at_1
749
- value: 0.0
750
  - type: mrr_at_10
751
- value: 0.0
752
  - type: mrr_at_100
753
- value: 0.0
754
  - type: mrr_at_1000
755
- value: 0.0
756
  - type: mrr_at_3
757
- value: 0.0
758
  - type: mrr_at_5
759
- value: 0.0
760
  - type: ndcg_at_1
761
- value: 0.0
762
  - type: ndcg_at_10
763
  value: 45.629999999999995
764
  - type: ndcg_at_100
765
- value: 0.0
766
  - type: ndcg_at_1000
767
- value: 0.0
768
  - type: ndcg_at_3
769
- value: 0.0
770
  - type: ndcg_at_5
771
- value: 0.0
772
  - type: precision_at_1
773
- value: 0.0
774
  - type: precision_at_10
775
  value: 7.119000000000001
776
  - type: precision_at_100
777
- value: 0.0
778
  - type: precision_at_1000
779
- value: 0.0
780
  - type: precision_at_3
781
- value: 0.0
782
  - type: precision_at_5
783
- value: 0.0
784
  - type: recall_at_1
785
- value: 0.0
786
  - type: recall_at_10
787
  value: 67.972
788
  - type: recall_at_100
789
- value: 0.0
790
  - type: recall_at_1000
791
- value: 0.0
792
  - type: recall_at_3
793
- value: 0.0
794
  - type: recall_at_5
795
- value: 0.0
796
  - type: main_score
797
  value: 45.629999999999995
798
  task:
@@ -938,17 +939,17 @@ model-index:
938
  - type: map_at_5
939
  value: 15.171000000000001
940
  - type: mrr_at_1
941
- value: 0.0
942
  - type: mrr_at_10
943
- value: 0.0
944
  - type: mrr_at_100
945
- value: 0.0
946
  - type: mrr_at_1000
947
- value: 0.0
948
  - type: mrr_at_3
949
- value: 0.0
950
  - type: mrr_at_5
951
- value: 0.0
952
  - type: ndcg_at_1
953
  value: 55.728
954
  - type: ndcg_at_10
@@ -1009,17 +1010,17 @@ model-index:
1009
  - type: map_at_5
1010
  value: 65.364
1011
  - type: mrr_at_1
1012
- value: 0.0
1013
  - type: mrr_at_10
1014
- value: 0.0
1015
  - type: mrr_at_100
1016
- value: 0.0
1017
  - type: mrr_at_1000
1018
- value: 0.0
1019
  - type: mrr_at_3
1020
- value: 0.0
1021
  - type: mrr_at_5
1022
- value: 0.0
1023
  - type: ndcg_at_1
1024
  value: 55.417
1025
  - type: ndcg_at_10
@@ -1080,17 +1081,17 @@ model-index:
1080
  - type: map_at_5
1081
  value: 84.396
1082
  - type: mrr_at_1
1083
- value: 0.0
1084
  - type: mrr_at_10
1085
- value: 0.0
1086
  - type: mrr_at_100
1087
- value: 0.0
1088
  - type: mrr_at_1000
1089
- value: 0.0
1090
  - type: mrr_at_3
1091
- value: 0.0
1092
  - type: mrr_at_5
1093
- value: 0.0
1094
  - type: ndcg_at_1
1095
  value: 82.12
1096
  - type: ndcg_at_10
@@ -1181,17 +1182,17 @@ model-index:
1181
  - type: map_at_5
1182
  value: 11.158
1183
  - type: mrr_at_1
1184
- value: 0.0
1185
  - type: mrr_at_10
1186
- value: 0.0
1187
  - type: mrr_at_100
1188
- value: 0.0
1189
  - type: mrr_at_1000
1190
- value: 0.0
1191
  - type: mrr_at_3
1192
- value: 0.0
1193
  - type: mrr_at_5
1194
- value: 0.0
1195
  - type: ndcg_at_1
1196
  value: 26.3
1197
  - type: ndcg_at_10
@@ -1474,19 +1475,19 @@ model-index:
1474
  - type: map_at_5
1475
  value: 74.74
1476
  - type: mrr_at_1
1477
- value: 0.0
1478
  - type: mrr_at_10
1479
- value: 0.0
1480
  - type: mrr_at_100
1481
- value: 0.0
1482
  - type: mrr_at_1000
1483
- value: 0.0
1484
  - type: mrr_at_3
1485
- value: 0.0
1486
  - type: mrr_at_5
1487
- value: 0.0
1488
  - type: ndcg_at_1
1489
- value: 66.0
1490
  - type: ndcg_at_10
1491
  value: 80.12700000000001
1492
  - type: ndcg_at_100
@@ -1498,7 +1499,7 @@ model-index:
1498
  - type: ndcg_at_5
1499
  value: 78.827
1500
  - type: precision_at_1
1501
- value: 66.0
1502
  - type: precision_at_10
1503
  value: 10.567
1504
  - type: precision_at_100
@@ -1516,7 +1517,7 @@ model-index:
1516
  - type: recall_at_100
1517
  value: 98.667
1518
  - type: recall_at_1000
1519
- value: 100.0
1520
  - type: recall_at_3
1521
  value: 83.322
1522
  - type: recall_at_5
@@ -1680,19 +1681,19 @@ model-index:
1680
  - type: map_at_5
1681
  value: 1.185
1682
  - type: mrr_at_1
1683
- value: 0.0
1684
  - type: mrr_at_10
1685
- value: 0.0
1686
  - type: mrr_at_100
1687
- value: 0.0
1688
  - type: mrr_at_1000
1689
- value: 0.0
1690
  - type: mrr_at_3
1691
- value: 0.0
1692
  - type: mrr_at_5
1693
- value: 0.0
1694
  - type: ndcg_at_1
1695
- value: 91.0
1696
  - type: ndcg_at_10
1697
  value: 88.442
1698
  - type: ndcg_at_100
@@ -1704,7 +1705,7 @@ model-index:
1704
  - type: ndcg_at_5
1705
  value: 89.562
1706
  - type: precision_at_1
1707
- value: 92.0
1708
  - type: precision_at_10
1709
  value: 92.60000000000001
1710
  - type: precision_at_100
@@ -1712,7 +1713,7 @@ model-index:
1712
  - type: precision_at_1000
1713
  value: 28.222
1714
  - type: precision_at_3
1715
- value: 94.0
1716
  - type: precision_at_5
1717
  value: 93.60000000000001
1718
  - type: recall_at_1
@@ -1751,17 +1752,17 @@ model-index:
1751
  - type: map_at_5
1752
  value: 9.49
1753
  - type: mrr_at_1
1754
- value: 0.0
1755
  - type: mrr_at_10
1756
- value: 0.0
1757
  - type: mrr_at_100
1758
- value: 0.0
1759
  - type: mrr_at_1000
1760
- value: 0.0
1761
  - type: mrr_at_3
1762
- value: 0.0
1763
  - type: mrr_at_5
1764
- value: 0.0
1765
  - type: ndcg_at_1
1766
  value: 47.959
1767
  - type: ndcg_at_10
@@ -2004,7 +2005,7 @@ model-index:
2004
  language:
2005
  - en
2006
  license: cc-by-nc-4.0
2007
- library_name: transformers
2008
  ---
2009
  ## Introduction
2010
  We present NV-Embed-v2, a generalist embedding model that ranks No. 1 on the Massive Text Embedding Benchmark ([MTEB benchmark](https://huggingface.co/spaces/mteb/leaderboard))(as of Aug 30, 2024) with a score of 72.31 across 56 text embedding tasks. It also holds the No. 1 in the retrieval sub-category (a score of 62.65 across 15 tasks) in the leaderboard, which is essential to the development of RAG technology.
@@ -2049,7 +2050,7 @@ passages = [
2049
  model = AutoModel.from_pretrained('nvidia/NV-Embed-v2', trust_remote_code=True)
2050
 
2051
  # get the embeddings
2052
- max_length = 32768
2053
  query_embeddings = model.encode(queries, instruction=query_prefix, max_length=max_length)
2054
  passage_embeddings = model.encode(passages, instruction=passage_prefix, max_length=max_length)
2055
 
@@ -2064,7 +2065,7 @@ passage_embeddings = F.normalize(passage_embeddings, p=2, dim=1)
2064
 
2065
  scores = (query_embeddings @ passage_embeddings.T) * 100
2066
  print(scores.tolist())
2067
- # [[87.42693328857422, 0.46283677220344543], [0.965264618396759, 86.03721618652344]]
2068
  ```
2069
 
2070
 
@@ -2091,7 +2092,7 @@ passages = [
2091
 
2092
  # load model with tokenizer
2093
  model = SentenceTransformer('nvidia/NV-Embed-v2', trust_remote_code=True)
2094
- model.max_seq_length = 32768
2095
  model.tokenizer.padding_side="right"
2096
 
2097
  def add_eos(input_examples):
@@ -2114,7 +2115,7 @@ For commercial purpose, we recommend you to use the models of [NeMo Retriever Mi
2114
 
2115
 
2116
  ## Correspondence to
2117
- Chankyu Lee ([email protected]), Rajarshi Roy ([email protected]), Wei Ping ([email protected])
2118
 
2119
 
2120
  ## Citation
@@ -2185,4 +2186,4 @@ cd sentence-transformers
2185
  git checkout v2.7-release
2186
  # Modify L353 in SentenceTransformer.py to **'extra_features["prompt_length"] = tokenized_prompt["input_ids"].shape[-1]'**.
2187
  pip install -e .
2188
- ```
 
1
  ---
2
  tags:
3
  - mteb
4
+ - transformers
5
  - sentence-transformers
6
  model-index:
7
  - name: NV-Embed-v2
 
91
  - type: map_at_5
92
  value: 61.027
93
  - type: mrr_at_1
94
+ value: 0
95
  - type: mrr_at_10
96
+ value: 0
97
  - type: mrr_at_100
98
+ value: 0
99
  - type: mrr_at_1000
100
+ value: 0
101
  - type: mrr_at_3
102
+ value: 0
103
  - type: mrr_at_5
104
+ value: 0
105
  - type: ndcg_at_1
106
  value: 46.515
107
  - type: ndcg_at_10
 
279
  - type: map_at_5
280
  value: 42.317083333333336
281
  - type: mrr_at_1
282
+ value: 0
283
  - type: mrr_at_10
284
+ value: 0
285
  - type: mrr_at_100
286
+ value: 0
287
  - type: mrr_at_1000
288
+ value: 0
289
  - type: mrr_at_3
290
+ value: 0
291
  - type: mrr_at_5
292
+ value: 0
293
  - type: ndcg_at_1
294
  value: 38.30616666666667
295
  - type: ndcg_at_10
 
350
  - type: map_at_5
351
  value: 31.955
352
  - type: mrr_at_1
353
+ value: 0
354
  - type: mrr_at_10
355
+ value: 0
356
  - type: mrr_at_100
357
+ value: 0
358
  - type: mrr_at_1000
359
+ value: 0
360
  - type: mrr_at_3
361
+ value: 0
362
  - type: mrr_at_5
363
+ value: 0
364
  - type: ndcg_at_1
365
  value: 44.104
366
  - type: ndcg_at_10
 
421
  - type: map_at_5
422
  value: 21.062
423
  - type: mrr_at_1
424
+ value: 0
425
  - type: mrr_at_10
426
+ value: 0
427
  - type: mrr_at_100
428
+ value: 0
429
  - type: mrr_at_1000
430
+ value: 0
431
  - type: mrr_at_3
432
+ value: 0
433
  - type: mrr_at_5
434
+ value: 0
435
  - type: ndcg_at_1
436
+ value: 66
437
  - type: ndcg_at_10
438
  value: 53.496
439
  - type: ndcg_at_100
 
453
  - type: precision_at_1000
454
  value: 2.5940000000000003
455
  - type: precision_at_3
456
+ value: 61
457
  - type: precision_at_5
458
  value: 54.65
459
  - type: recall_at_1
 
511
  - type: map_at_5
512
  value: 91.262
513
  - type: mrr_at_1
514
+ value: 0
515
  - type: mrr_at_10
516
+ value: 0
517
  - type: mrr_at_100
518
+ value: 0
519
  - type: mrr_at_1000
520
+ value: 0
521
  - type: mrr_at_3
522
+ value: 0
523
  - type: mrr_at_5
524
+ value: 0
525
  - type: ndcg_at_1
526
  value: 91.20899999999999
527
  - type: ndcg_at_10
 
582
  - type: map_at_5
583
  value: 55.054
584
  - type: mrr_at_1
585
+ value: 0
586
  - type: mrr_at_10
587
+ value: 0
588
  - type: mrr_at_100
589
+ value: 0
590
  - type: mrr_at_1000
591
+ value: 0
592
  - type: mrr_at_3
593
+ value: 0
594
  - type: mrr_at_5
595
+ value: 0
596
  - type: ndcg_at_1
597
  value: 64.815
598
  - type: ndcg_at_10
 
653
  - type: map_at_5
654
  value: 78.935
655
  - type: mrr_at_1
656
+ value: 0
657
  - type: mrr_at_10
658
+ value: 0
659
  - type: mrr_at_100
660
+ value: 0
661
  - type: mrr_at_1000
662
+ value: 0
663
  - type: mrr_at_3
664
+ value: 0
665
  - type: mrr_at_5
666
+ value: 0
667
  - type: ndcg_at_1
668
  value: 89.305
669
  - type: ndcg_at_10
 
735
  type: mteb/msmarco
736
  metrics:
737
  - type: map_at_1
738
+ value: 0
739
  - type: map_at_10
740
  value: 38.342
741
  - type: map_at_100
742
+ value: 0
743
  - type: map_at_1000
744
+ value: 0
745
  - type: map_at_3
746
+ value: 0
747
  - type: map_at_5
748
+ value: 0
749
  - type: mrr_at_1
750
+ value: 0
751
  - type: mrr_at_10
752
+ value: 0
753
  - type: mrr_at_100
754
+ value: 0
755
  - type: mrr_at_1000
756
+ value: 0
757
  - type: mrr_at_3
758
+ value: 0
759
  - type: mrr_at_5
760
+ value: 0
761
  - type: ndcg_at_1
762
+ value: 0
763
  - type: ndcg_at_10
764
  value: 45.629999999999995
765
  - type: ndcg_at_100
766
+ value: 0
767
  - type: ndcg_at_1000
768
+ value: 0
769
  - type: ndcg_at_3
770
+ value: 0
771
  - type: ndcg_at_5
772
+ value: 0
773
  - type: precision_at_1
774
+ value: 0
775
  - type: precision_at_10
776
  value: 7.119000000000001
777
  - type: precision_at_100
778
+ value: 0
779
  - type: precision_at_1000
780
+ value: 0
781
  - type: precision_at_3
782
+ value: 0
783
  - type: precision_at_5
784
+ value: 0
785
  - type: recall_at_1
786
+ value: 0
787
  - type: recall_at_10
788
  value: 67.972
789
  - type: recall_at_100
790
+ value: 0
791
  - type: recall_at_1000
792
+ value: 0
793
  - type: recall_at_3
794
+ value: 0
795
  - type: recall_at_5
796
+ value: 0
797
  - type: main_score
798
  value: 45.629999999999995
799
  task:
 
939
  - type: map_at_5
940
  value: 15.171000000000001
941
  - type: mrr_at_1
942
+ value: 0
943
  - type: mrr_at_10
944
+ value: 0
945
  - type: mrr_at_100
946
+ value: 0
947
  - type: mrr_at_1000
948
+ value: 0
949
  - type: mrr_at_3
950
+ value: 0
951
  - type: mrr_at_5
952
+ value: 0
953
  - type: ndcg_at_1
954
  value: 55.728
955
  - type: ndcg_at_10
 
1010
  - type: map_at_5
1011
  value: 65.364
1012
  - type: mrr_at_1
1013
+ value: 0
1014
  - type: mrr_at_10
1015
+ value: 0
1016
  - type: mrr_at_100
1017
+ value: 0
1018
  - type: mrr_at_1000
1019
+ value: 0
1020
  - type: mrr_at_3
1021
+ value: 0
1022
  - type: mrr_at_5
1023
+ value: 0
1024
  - type: ndcg_at_1
1025
  value: 55.417
1026
  - type: ndcg_at_10
 
1081
  - type: map_at_5
1082
  value: 84.396
1083
  - type: mrr_at_1
1084
+ value: 0
1085
  - type: mrr_at_10
1086
+ value: 0
1087
  - type: mrr_at_100
1088
+ value: 0
1089
  - type: mrr_at_1000
1090
+ value: 0
1091
  - type: mrr_at_3
1092
+ value: 0
1093
  - type: mrr_at_5
1094
+ value: 0
1095
  - type: ndcg_at_1
1096
  value: 82.12
1097
  - type: ndcg_at_10
 
1182
  - type: map_at_5
1183
  value: 11.158
1184
  - type: mrr_at_1
1185
+ value: 0
1186
  - type: mrr_at_10
1187
+ value: 0
1188
  - type: mrr_at_100
1189
+ value: 0
1190
  - type: mrr_at_1000
1191
+ value: 0
1192
  - type: mrr_at_3
1193
+ value: 0
1194
  - type: mrr_at_5
1195
+ value: 0
1196
  - type: ndcg_at_1
1197
  value: 26.3
1198
  - type: ndcg_at_10
 
1475
  - type: map_at_5
1476
  value: 74.74
1477
  - type: mrr_at_1
1478
+ value: 0
1479
  - type: mrr_at_10
1480
+ value: 0
1481
  - type: mrr_at_100
1482
+ value: 0
1483
  - type: mrr_at_1000
1484
+ value: 0
1485
  - type: mrr_at_3
1486
+ value: 0
1487
  - type: mrr_at_5
1488
+ value: 0
1489
  - type: ndcg_at_1
1490
+ value: 66
1491
  - type: ndcg_at_10
1492
  value: 80.12700000000001
1493
  - type: ndcg_at_100
 
1499
  - type: ndcg_at_5
1500
  value: 78.827
1501
  - type: precision_at_1
1502
+ value: 66
1503
  - type: precision_at_10
1504
  value: 10.567
1505
  - type: precision_at_100
 
1517
  - type: recall_at_100
1518
  value: 98.667
1519
  - type: recall_at_1000
1520
+ value: 100
1521
  - type: recall_at_3
1522
  value: 83.322
1523
  - type: recall_at_5
 
1681
  - type: map_at_5
1682
  value: 1.185
1683
  - type: mrr_at_1
1684
+ value: 0
1685
  - type: mrr_at_10
1686
+ value: 0
1687
  - type: mrr_at_100
1688
+ value: 0
1689
  - type: mrr_at_1000
1690
+ value: 0
1691
  - type: mrr_at_3
1692
+ value: 0
1693
  - type: mrr_at_5
1694
+ value: 0
1695
  - type: ndcg_at_1
1696
+ value: 91
1697
  - type: ndcg_at_10
1698
  value: 88.442
1699
  - type: ndcg_at_100
 
1705
  - type: ndcg_at_5
1706
  value: 89.562
1707
  - type: precision_at_1
1708
+ value: 92
1709
  - type: precision_at_10
1710
  value: 92.60000000000001
1711
  - type: precision_at_100
 
1713
  - type: precision_at_1000
1714
  value: 28.222
1715
  - type: precision_at_3
1716
+ value: 94
1717
  - type: precision_at_5
1718
  value: 93.60000000000001
1719
  - type: recall_at_1
 
1752
  - type: map_at_5
1753
  value: 9.49
1754
  - type: mrr_at_1
1755
+ value: 0
1756
  - type: mrr_at_10
1757
+ value: 0
1758
  - type: mrr_at_100
1759
+ value: 0
1760
  - type: mrr_at_1000
1761
+ value: 0
1762
  - type: mrr_at_3
1763
+ value: 0
1764
  - type: mrr_at_5
1765
+ value: 0
1766
  - type: ndcg_at_1
1767
  value: 47.959
1768
  - type: ndcg_at_10
 
2005
  language:
2006
  - en
2007
  license: cc-by-nc-4.0
2008
+ base_model: mistralai/Mistral-7B-v0.1
2009
  ---
2010
  ## Introduction
2011
  We present NV-Embed-v2, a generalist embedding model that ranks No. 1 on the Massive Text Embedding Benchmark ([MTEB benchmark](https://huggingface.co/spaces/mteb/leaderboard))(as of Aug 30, 2024) with a score of 72.31 across 56 text embedding tasks. It also holds the No. 1 in the retrieval sub-category (a score of 62.65 across 15 tasks) in the leaderboard, which is essential to the development of RAG technology.
 
2050
  model = AutoModel.from_pretrained('nvidia/NV-Embed-v2', trust_remote_code=True)
2051
 
2052
  # get the embeddings
2053
+ max_length = 4096
2054
  query_embeddings = model.encode(queries, instruction=query_prefix, max_length=max_length)
2055
  passage_embeddings = model.encode(passages, instruction=passage_prefix, max_length=max_length)
2056
 
 
2065
 
2066
  scores = (query_embeddings @ passage_embeddings.T) * 100
2067
  print(scores.tolist())
2068
+ # [[87.42692565917969, 0.462837278842926], [0.9652643203735352, 86.0372314453125]]
2069
  ```
2070
 
2071
 
 
2092
 
2093
  # load model with tokenizer
2094
  model = SentenceTransformer('nvidia/NV-Embed-v2', trust_remote_code=True)
2095
+ model.max_seq_length = 4096
2096
  model.tokenizer.padding_side="right"
2097
 
2098
  def add_eos(input_examples):
 
2115
 
2116
 
2117
  ## Correspondence to
2118
+ Chankyu Lee ([email protected]), Wei Ping ([email protected])
2119
 
2120
 
2121
  ## Citation
 
2186
  git checkout v2.7-release
2187
  # Modify L353 in SentenceTransformer.py to **'extra_features["prompt_length"] = tokenized_prompt["input_ids"].shape[-1]'**.
2188
  pip install -e .
2189
+ ```
config.json CHANGED
@@ -1,5 +1,5 @@
1
  {
2
- "_name_or_path": "nvidia/NV-Embed-v2",
3
  "add_eos": true,
4
  "add_pad_token": true,
5
  "architectures": [
@@ -18,7 +18,7 @@
18
  "model_type": "nvembed",
19
  "padding_side": "right",
20
  "text_config": {
21
- "_name_or_path": "nvidia/NV-Embed-v2",
22
  "add_cross_attention": false,
23
  "architectures": [
24
  "MistralModel"
 
1
  {
2
+ "_name_or_path": "nvidia/NV-Embed-v1",
3
  "add_eos": true,
4
  "add_pad_token": true,
5
  "architectures": [
 
18
  "model_type": "nvembed",
19
  "padding_side": "right",
20
  "text_config": {
21
+ "_name_or_path": "nvidia/NV-Embed-v1",
22
  "add_cross_attention": false,
23
  "architectures": [
24
  "MistralModel"
configuration_nvembed.py CHANGED
@@ -76,8 +76,6 @@ class LatentAttentionConfig(PretrainedConfig):
76
  self.latent_dim = latent_dim
77
  self.cross_dim_head = cross_dim_head
78
 
79
- super().__init__(**kwargs)
80
-
81
 
82
  class BidirectionalMistralConfig(MistralConfig):
83
  model_type = BIDIR_MISTRAL_TYPE
 
76
  self.latent_dim = latent_dim
77
  self.cross_dim_head = cross_dim_head
78
 
 
 
79
 
80
  class BidirectionalMistralConfig(MistralConfig):
81
  model_type = BIDIR_MISTRAL_TYPE