Hugging Face's logo

Spaces:
uonlp
/
open_multilingual_llm_leaderboard
Running

App Files Community
8
laiviet commited on
Commit
95b0e17
•
1 Parent(s): f067bfb

Update evals

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. evals/arc-challenge/arc_ar_challenge_bloom-1b7.json +0 -23
  2. evals/arc-challenge/arc_ar_challenge_bloom-560.json +0 -23
  3. evals/arc-challenge/arc_ar_challenge_bloom-7b1.json +0 -23
  4. evals/arc-challenge/arc_ar_challenge_gpt2-large.json +0 -23
  5. evals/arc-challenge/arc_ar_challenge_gpt2-medium.json +0 -23
  6. evals/arc-challenge/arc_ar_challenge_gpt2.json +0 -23
  7. evals/arc-challenge/arc_ar_challenge_llama-7B.json +0 -23
  8. evals/arc-challenge/arc_bn_challenge_bloom-1b7.json +0 -23
  9. evals/arc-challenge/arc_bn_challenge_bloom-560.json +0 -23
  10. evals/arc-challenge/arc_bn_challenge_bloom-7b1.json +0 -23
  11. evals/arc-challenge/arc_bn_challenge_gpt2-medium.json +0 -23
  12. evals/arc-challenge/arc_bn_challenge_gpt2.json +0 -23
  13. evals/arc-challenge/arc_bn_challenge_llama-7B.json +0 -23
  14. evals/arc-challenge/arc_ca_challenge_bloom-1b7.json +0 -23
  15. evals/arc-challenge/arc_ca_challenge_bloom-560.json +0 -23
  16. evals/arc-challenge/arc_ca_challenge_bloom-7b1.json +0 -23
  17. evals/arc-challenge/arc_ca_challenge_gpt2-large.json +0 -23
  18. evals/arc-challenge/arc_ca_challenge_gpt2-medium.json +0 -23
  19. evals/arc-challenge/arc_ca_challenge_gpt2.json +0 -23
  20. evals/arc-challenge/arc_ca_challenge_llama-7B.json +0 -23
  21. evals/arc-challenge/arc_da_challenge_bloom-1b7.json +0 -23
  22. evals/arc-challenge/arc_da_challenge_bloom-560.json +0 -23
  23. evals/arc-challenge/arc_da_challenge_bloom-7b1.json +0 -23
  24. evals/arc-challenge/arc_da_challenge_gpt2-large.json +0 -23
  25. evals/arc-challenge/arc_da_challenge_gpt2-medium.json +0 -23
  26. evals/arc-challenge/arc_da_challenge_llama-7B.json +0 -23
  27. evals/arc-challenge/arc_de_challenge_bloom-1b7.json +0 -23
  28. evals/arc-challenge/arc_de_challenge_bloom-560.json +0 -23
  29. evals/arc-challenge/arc_de_challenge_bloom-7b1.json +0 -23
  30. evals/arc-challenge/arc_de_challenge_gpt2-large.json +0 -23
  31. evals/arc-challenge/arc_de_challenge_gpt2-medium.json +0 -23
  32. evals/arc-challenge/arc_de_challenge_gpt2.json +0 -23
  33. evals/arc-challenge/arc_de_challenge_llama-7B.json +0 -23
  34. evals/arc-challenge/arc_es_challenge_bloom-1b7.json +0 -23
  35. evals/arc-challenge/arc_es_challenge_bloom-560.json +0 -23
  36. evals/arc-challenge/arc_es_challenge_bloom-7b1.json +0 -23
  37. evals/arc-challenge/arc_es_challenge_gpt2-large.json +0 -23
  38. evals/arc-challenge/arc_es_challenge_gpt2-medium.json +0 -23
  39. evals/arc-challenge/arc_es_challenge_gpt2.json +0 -23
  40. evals/arc-challenge/arc_es_challenge_llama-7B.json +0 -23
  41. evals/arc-challenge/arc_eu_challenge_bloom-1b7.json +0 -23
  42. evals/arc-challenge/arc_eu_challenge_bloom-560.json +0 -23
  43. evals/arc-challenge/arc_eu_challenge_bloom-7b1.json +0 -23
  44. evals/arc-challenge/arc_eu_challenge_gpt2-large.json +0 -23
  45. evals/arc-challenge/arc_eu_challenge_gpt2-medium.json +0 -23
  46. evals/arc-challenge/arc_eu_challenge_gpt2.json +0 -23
  47. evals/arc-challenge/arc_eu_challenge_llama-7B.json +0 -23
  48. evals/arc-challenge/arc_fr_challenge_bloom-1b7.json +0 -23
  49. evals/arc-challenge/arc_fr_challenge_bloom-560.json +0 -23
  50. evals/arc-challenge/arc_fr_challenge_bloom-7b1.json +0 -23
evals/arc-challenge/arc_ar_challenge_bloom-1b7.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_ar_challenge": {
4
- "acc": 0.22818791946308725,
5
- "acc_stderr": 0.02435139725761051,
6
- "acc_norm": 0.2516778523489933,
7
- "acc_norm_stderr": 0.025181904610615872
8
- }
9
- },
10
- "versions": {
11
- "arc_ar_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-1b7",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_ar_challenge_bloom-560.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_ar_challenge": {
4
- "acc": 0.2550335570469799,
5
- "acc_stderr": 0.025292327380712708,
6
- "acc_norm": 0.2550335570469799,
7
- "acc_norm_stderr": 0.025292327380712708
8
- }
9
- },
10
- "versions": {
11
- "arc_ar_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-560m",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_ar_challenge_bloom-7b1.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_ar_challenge": {
4
- "acc": 0.28187919463087246,
5
- "acc_stderr": 0.026106703750007426,
6
- "acc_norm": 0.3087248322147651,
7
- "acc_norm_stderr": 0.026806063072940547
8
- }
9
- },
10
- "versions": {
11
- "arc_ar_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/bloom-7b1",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_ar_challenge_gpt2-large.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_ar_challenge": {
4
- "acc": 0.20134228187919462,
5
- "acc_stderr": 0.023268565767685306,
6
- "acc_norm": 0.21476510067114093,
7
- "acc_norm_stderr": 0.023828868848284352
8
- }
9
- },
10
- "versions": {
11
- "arc_ar_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2-large",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_ar_challenge_gpt2-medium.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_ar_challenge": {
4
- "acc": 0.19463087248322147,
5
- "acc_stderr": 0.022973392306598162,
6
- "acc_norm": 0.21140939597315436,
7
- "acc_norm_stderr": 0.02369243605357901
8
- }
9
- },
10
- "versions": {
11
- "arc_ar_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2-medium",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_ar_challenge_gpt2.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_ar_challenge": {
4
- "acc": 0.20134228187919462,
5
- "acc_stderr": 0.023268565767685313,
6
- "acc_norm": 0.22483221476510068,
7
- "acc_norm_stderr": 0.024224169829650755
8
- }
9
- },
10
- "versions": {
11
- "arc_ar_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_ar_challenge_llama-7B.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_ar_challenge": {
4
- "acc": 0.22483221476510068,
5
- "acc_stderr": 0.02422416982965075,
6
- "acc_norm": 0.24161073825503357,
7
- "acc_norm_stderr": 0.024838535108028477
8
- }
9
- },
10
- "versions": {
11
- "arc_ar_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/llama-7B",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_bn_challenge_bloom-1b7.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_bn_challenge": {
4
- "acc": 0.20945945945945946,
5
- "acc_stderr": 0.023691963473475724,
6
- "acc_norm": 0.2533783783783784,
7
- "acc_norm_stderr": 0.025323518629100008
8
- }
9
- },
10
- "versions": {
11
- "arc_bn_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-1b7",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_bn_challenge_bloom-560.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_bn_challenge": {
4
- "acc": 0.22972972972972974,
5
- "acc_stderr": 0.024491712953916975,
6
- "acc_norm": 0.24662162162162163,
7
- "acc_norm_stderr": 0.025096383517594287
8
- }
9
- },
10
- "versions": {
11
- "arc_bn_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-560m",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_bn_challenge_bloom-7b1.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_bn_challenge": {
4
- "acc": 0.23986486486486486,
5
- "acc_stderr": 0.02486094967084638,
6
- "acc_norm": 0.28040540540540543,
7
- "acc_norm_stderr": 0.026153277917823237
8
- }
9
- },
10
- "versions": {
11
- "arc_bn_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/bloom-7b1",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_bn_challenge_gpt2-medium.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_bn_challenge": {
4
- "acc": 0.20608108108108109,
5
- "acc_stderr": 0.02355028295929425,
6
- "acc_norm": 0.24662162162162163,
7
- "acc_norm_stderr": 0.02509638351759427
8
- }
9
- },
10
- "versions": {
11
- "arc_bn_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2-medium",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_bn_challenge_gpt2.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_bn_challenge": {
4
- "acc": 0.22635135135135134,
5
- "acc_stderr": 0.024364215012920555,
6
- "acc_norm": 0.2668918918918919,
7
- "acc_norm_stderr": 0.025753762926257917
8
- }
9
- },
10
- "versions": {
11
- "arc_bn_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_bn_challenge_llama-7B.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_bn_challenge": {
4
- "acc": 0.22635135135135134,
5
- "acc_stderr": 0.024364215012920565,
6
- "acc_norm": 0.26013513513513514,
7
- "acc_norm_stderr": 0.02554257639364025
8
- }
9
- },
10
- "versions": {
11
- "arc_bn_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/llama-7B",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_ca_challenge_bloom-1b7.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_ca_challenge": {
4
- "acc": 0.2356902356902357,
5
- "acc_stderr": 0.02466946003490763,
6
- "acc_norm": 0.27946127946127947,
7
- "acc_norm_stderr": 0.026082164400369843
8
- }
9
- },
10
- "versions": {
11
- "arc_ca_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-1b7",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_ca_challenge_bloom-560.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_ca_challenge": {
4
- "acc": 0.2053872053872054,
5
- "acc_stderr": 0.02348110951859932,
6
- "acc_norm": 0.23232323232323232,
7
- "acc_norm_stderr": 0.02454650495612789
8
- }
9
- },
10
- "versions": {
11
- "arc_ca_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-560m",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_ca_challenge_bloom-7b1.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_ca_challenge": {
4
- "acc": 0.3164983164983165,
5
- "acc_stderr": 0.02703395838420779,
6
- "acc_norm": 0.3434343434343434,
7
- "acc_norm_stderr": 0.0276003816062635
8
- }
9
- },
10
- "versions": {
11
- "arc_ca_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/bloom-7b1",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_ca_challenge_gpt2-large.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_ca_challenge": {
4
- "acc": 0.20875420875420875,
5
- "acc_stderr": 0.02362258775627148,
6
- "acc_norm": 0.22895622895622897,
7
- "acc_norm_stderr": 0.02442136264227106
8
- }
9
- },
10
- "versions": {
11
- "arc_ca_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2-large",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_ca_challenge_gpt2-medium.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_ca_challenge": {
4
- "acc": 0.20875420875420875,
5
- "acc_stderr": 0.023622587756271473,
6
- "acc_norm": 0.21212121212121213,
7
- "acc_norm_stderr": 0.023761611918761673
8
- }
9
- },
10
- "versions": {
11
- "arc_ca_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2-medium",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_ca_challenge_gpt2.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_ca_challenge": {
4
- "acc": 0.21885521885521886,
5
- "acc_stderr": 0.024032467624412215,
6
- "acc_norm": 0.21885521885521886,
7
- "acc_norm_stderr": 0.02403246762441221
8
- }
9
- },
10
- "versions": {
11
- "arc_ca_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_ca_challenge_llama-7B.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_ca_challenge": {
4
- "acc": 0.29292929292929293,
5
- "acc_stderr": 0.026452514969665927,
6
- "acc_norm": 0.29292929292929293,
7
- "acc_norm_stderr": 0.02645251496966592
8
- }
9
- },
10
- "versions": {
11
- "arc_ca_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/llama-7B",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_da_challenge_bloom-1b7.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_da_challenge": {
4
- "acc": 0.2255892255892256,
5
- "acc_stderr": 0.02429399929295737,
6
- "acc_norm": 0.26262626262626265,
7
- "acc_norm_stderr": 0.02557802773320011
8
- }
9
- },
10
- "versions": {
11
- "arc_da_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-1b7",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_da_challenge_bloom-560.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_da_challenge": {
4
- "acc": 0.25925925925925924,
5
- "acc_stderr": 0.025471492792791667,
6
- "acc_norm": 0.24579124579124578,
7
- "acc_norm_stderr": 0.025025521384235284
8
- }
9
- },
10
- "versions": {
11
- "arc_da_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-560m",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_da_challenge_bloom-7b1.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_da_challenge": {
4
- "acc": 0.24242424242424243,
5
- "acc_stderr": 0.02490893747050877,
6
- "acc_norm": 0.24915824915824916,
7
- "acc_norm_stderr": 0.025140041284626418
8
- }
9
- },
10
- "versions": {
11
- "arc_da_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/bloom-7b1",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_da_challenge_gpt2-large.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_da_challenge": {
4
- "acc": 0.23232323232323232,
5
- "acc_stderr": 0.02454650495612789,
6
- "acc_norm": 0.24242424242424243,
7
- "acc_norm_stderr": 0.024908937470508753
8
- }
9
- },
10
- "versions": {
11
- "arc_da_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2-large",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_da_challenge_gpt2-medium.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_da_challenge": {
4
- "acc": 0.24579124579124578,
5
- "acc_stderr": 0.0250255213842353,
6
- "acc_norm": 0.2727272727272727,
7
- "acc_norm_stderr": 0.025886127156886297
8
- }
9
- },
10
- "versions": {
11
- "arc_da_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2-medium",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_da_challenge_llama-7B.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_da_challenge": {
4
- "acc": 0.3063973063973064,
5
- "acc_stderr": 0.026794891419479452,
6
- "acc_norm": 0.3367003367003367,
7
- "acc_norm_stderr": 0.02746823841289221
8
- }
9
- },
10
- "versions": {
11
- "arc_da_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/llama-7B",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_de_challenge_bloom-1b7.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_de_challenge": {
4
- "acc": 0.24496644295302014,
5
- "acc_stderr": 0.024955035980898946,
6
- "acc_norm": 0.2953020134228188,
7
- "acc_norm_stderr": 0.026470155629081085
8
- }
9
- },
10
- "versions": {
11
- "arc_de_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-1b7",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_de_challenge_bloom-560.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_de_challenge": {
4
- "acc": 0.2348993288590604,
5
- "acc_stderr": 0.024599255015999244,
6
- "acc_norm": 0.28187919463087246,
7
- "acc_norm_stderr": 0.026106703750007426
8
- }
9
- },
10
- "versions": {
11
- "arc_de_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-560m",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_de_challenge_bloom-7b1.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_de_challenge": {
4
- "acc": 0.2684563758389262,
5
- "acc_stderr": 0.0257145395148175,
6
- "acc_norm": 0.2684563758389262,
7
- "acc_norm_stderr": 0.0257145395148175
8
- }
9
- },
10
- "versions": {
11
- "arc_de_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/bloom-7b1",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_de_challenge_gpt2-large.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_de_challenge": {
4
- "acc": 0.23825503355704697,
5
- "acc_stderr": 0.024719951493159625,
6
- "acc_norm": 0.27181208053691275,
7
- "acc_norm_stderr": 0.025815342279487567
8
- }
9
- },
10
- "versions": {
11
- "arc_de_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2-large",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_de_challenge_gpt2-medium.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_de_challenge": {
4
- "acc": 0.23825503355704697,
5
- "acc_stderr": 0.024719951493159625,
6
- "acc_norm": 0.28859060402684567,
7
- "acc_norm_stderr": 0.026291942108676806
8
- }
9
- },
10
- "versions": {
11
- "arc_de_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2-medium",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_de_challenge_gpt2.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_de_challenge": {
4
- "acc": 0.22483221476510068,
5
- "acc_stderr": 0.02422416982965075,
6
- "acc_norm": 0.21140939597315436,
7
- "acc_norm_stderr": 0.02369243605357901
8
- }
9
- },
10
- "versions": {
11
- "arc_de_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_de_challenge_llama-7B.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_de_challenge": {
4
- "acc": 0.2785234899328859,
5
- "acc_stderr": 0.0260114035784859,
6
- "acc_norm": 0.348993288590604,
7
- "acc_norm_stderr": 0.027658144793750224
8
- }
9
- },
10
- "versions": {
11
- "arc_de_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/llama-7B",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_es_challenge_bloom-1b7.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_es_challenge": {
4
- "acc": 0.2356902356902357,
5
- "acc_stderr": 0.02466946003490763,
6
- "acc_norm": 0.2895622895622896,
7
- "acc_norm_stderr": 0.026362594432681956
8
- }
9
- },
10
- "versions": {
11
- "arc_es_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-1b7",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_es_challenge_bloom-560.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_es_challenge": {
4
- "acc": 0.2255892255892256,
5
- "acc_stderr": 0.024293999292957367,
6
- "acc_norm": 0.2356902356902357,
7
- "acc_norm_stderr": 0.02466946003490764
8
- }
9
- },
10
- "versions": {
11
- "arc_es_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-560m",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_es_challenge_bloom-7b1.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_es_challenge": {
4
- "acc": 0.3265993265993266,
5
- "acc_stderr": 0.027258287015652305,
6
- "acc_norm": 0.3602693602693603,
7
- "acc_norm_stderr": 0.02790399493827167
8
- }
9
- },
10
- "versions": {
11
- "arc_es_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/bloom-7b1",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_es_challenge_gpt2-large.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_es_challenge": {
4
- "acc": 0.2222222222222222,
5
- "acc_stderr": 0.024164379788935483,
6
- "acc_norm": 0.26262626262626265,
7
- "acc_norm_stderr": 0.02557802773320012
8
- }
9
- },
10
- "versions": {
11
- "arc_es_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2-large",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_es_challenge_gpt2-medium.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_es_challenge": {
4
- "acc": 0.1919191919191919,
5
- "acc_stderr": 0.022889733897083934,
6
- "acc_norm": 0.25252525252525254,
7
- "acc_norm_stderr": 0.02525252525252536
8
- }
9
- },
10
- "versions": {
11
- "arc_es_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2-medium",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_es_challenge_gpt2.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_es_challenge": {
4
- "acc": 0.19865319865319866,
5
- "acc_stderr": 0.023190610381322127,
6
- "acc_norm": 0.24579124579124578,
7
- "acc_norm_stderr": 0.0250255213842353
8
- }
9
- },
10
- "versions": {
11
- "arc_es_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_es_challenge_llama-7B.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_es_challenge": {
4
- "acc": 0.3501683501683502,
5
- "acc_stderr": 0.027726370308831506,
6
- "acc_norm": 0.3602693602693603,
7
- "acc_norm_stderr": 0.02790399493827167
8
- }
9
- },
10
- "versions": {
11
- "arc_es_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/llama-7B",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_eu_challenge_bloom-1b7.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_eu_challenge": {
4
- "acc": 0.22377622377622378,
5
- "acc_stderr": 0.02468755105337312,
6
- "acc_norm": 0.2517482517482518,
7
- "acc_norm_stderr": 0.02570896966075011
8
- }
9
- },
10
- "versions": {
11
- "arc_eu_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-1b7",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_eu_challenge_bloom-560.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_eu_challenge": {
4
- "acc": 0.24475524475524477,
5
- "acc_stderr": 0.02546756553847068,
6
- "acc_norm": 0.19230769230769232,
7
- "acc_norm_stderr": 0.023345268410264786
8
- }
9
- },
10
- "versions": {
11
- "arc_eu_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-560m",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_eu_challenge_bloom-7b1.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_eu_challenge": {
4
- "acc": 0.23076923076923078,
5
- "acc_stderr": 0.024957141712425013,
6
- "acc_norm": 0.24125874125874125,
7
- "acc_norm_stderr": 0.025343462496583764
8
- }
9
- },
10
- "versions": {
11
- "arc_eu_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/bloom-7b1",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_eu_challenge_gpt2-large.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_eu_challenge": {
4
- "acc": 0.25874125874125875,
5
- "acc_stderr": 0.02594151450124707,
6
- "acc_norm": 0.24125874125874125,
7
- "acc_norm_stderr": 0.025343462496583737
8
- }
9
- },
10
- "versions": {
11
- "arc_eu_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2-large",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_eu_challenge_gpt2-medium.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_eu_challenge": {
4
- "acc": 0.2762237762237762,
5
- "acc_stderr": 0.026485626798716442,
6
- "acc_norm": 0.25874125874125875,
7
- "acc_norm_stderr": 0.025941514501247064
8
- }
9
- },
10
- "versions": {
11
- "arc_eu_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2-medium",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_eu_challenge_gpt2.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_eu_challenge": {
4
- "acc": 0.2762237762237762,
5
- "acc_stderr": 0.026485626798716456,
6
- "acc_norm": 0.24825174825174826,
7
- "acc_norm_stderr": 0.025589390464738234
8
- }
9
- },
10
- "versions": {
11
- "arc_eu_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=gpt2",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_eu_challenge_llama-7B.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_eu_challenge": {
4
- "acc": 0.26223776223776224,
5
- "acc_stderr": 0.026054539173797044,
6
- "acc_norm": 0.23426573426573427,
7
- "acc_norm_stderr": 0.02508828621716978
8
- }
9
- },
10
- "versions": {
11
- "arc_eu_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/llama-7B",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_fr_challenge_bloom-1b7.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_fr_challenge": {
4
- "acc": 0.2550335570469799,
5
- "acc_stderr": 0.025292327380712687,
6
- "acc_norm": 0.2953020134228188,
7
- "acc_norm_stderr": 0.026470155629081078
8
- }
9
- },
10
- "versions": {
11
- "arc_fr_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-1b7",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_fr_challenge_bloom-560.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_fr_challenge": {
4
- "acc": 0.2348993288590604,
5
- "acc_stderr": 0.024599255015999244,
6
- "acc_norm": 0.25838926174496646,
7
- "acc_norm_stderr": 0.025400777524610105
8
- }
9
- },
10
- "versions": {
11
- "arc_fr_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=bigscience/bloom-560m",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
evals/arc-challenge/arc_fr_challenge_bloom-7b1.json DELETED
@@ -1,23 +0,0 @@
1
- {
2
- "results": {
3
- "arc_fr_challenge": {
4
- "acc": 0.36577181208053694,
5
- "acc_stderr": 0.027947930997299652,
6
- "acc_norm": 0.3825503355704698,
7
- "acc_norm_stderr": 0.02820115194087938
8
- }
9
- },
10
- "versions": {
11
- "arc_fr_challenge": 0
12
- },
13
- "config": {
14
- "model": "hf-auto",
15
- "model_args": "pretrained=/sensei-fs/users/daclai/uoChatGPT/bloom-7b1",
16
- "batch_size": "1",
17
- "device": "cuda",
18
- "no_cache": false,
19
- "limit": null,
20
- "bootstrap_iters": 100000,
21
- "description_dict": {}
22
- }
23
- }