Upload 9 files
Browse files- benchmark_scores/result_jaqket_v2-0.2-0.2.json +24 -0
- benchmark_scores/result_jcola.json +26 -0
- benchmark_scores/result_jcommonsenseqa-1.1-0.2.1.json +26 -0
- benchmark_scores/result_jnli-1.3-0.2.json +30 -0
- benchmark_scores/result_jsquad-1.2-0.2.json +24 -0
- benchmark_scores/result_marc_ja-1.1-0.2.json +30 -0
- benchmark_scores/result_mgsm.json +24 -0
- benchmark_scores/result_xlsum_ja.json +23 -0
- benchmark_scores/result_xwinograd_ja.json +24 -0
benchmark_scores/result_jaqket_v2-0.2-0.2.json
ADDED
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"jaqket_v2-0.2-0.2": {
|
4 |
+
"exact_match": 80.2405498281787,
|
5 |
+
"f1": 84.5903685852139
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"jaqket_v2-0.2-0.2": 0.2
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "hf-causal-experimental",
|
13 |
+
"model_args": "pretrained=HachiML/youri-2x7b_dev,tokenizer=HachiML/youri-2x7b_dev,use_accelerate=True,dtype=auto",
|
14 |
+
"num_fewshot": [
|
15 |
+
1
|
16 |
+
],
|
17 |
+
"batch_size": 2,
|
18 |
+
"device": "cuda",
|
19 |
+
"no_cache": false,
|
20 |
+
"limit": null,
|
21 |
+
"bootstrap_iters": 100000,
|
22 |
+
"description_dict": null
|
23 |
+
}
|
24 |
+
}
|
benchmark_scores/result_jcola.json
ADDED
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"jcola": {
|
4 |
+
"balanced_acc": 0.5909437739064947,
|
5 |
+
"mcc": 0.14568482514613612,
|
6 |
+
"mcc_stderr": 0.03657851423629269,
|
7 |
+
"macro_f1": 0.5553402665190399
|
8 |
+
}
|
9 |
+
},
|
10 |
+
"versions": {
|
11 |
+
"jcola": 0.2
|
12 |
+
},
|
13 |
+
"config": {
|
14 |
+
"model": "hf-causal-experimental",
|
15 |
+
"model_args": "pretrained=HachiML/youri-2x7b_dev,tokenizer=HachiML/youri-2x7b_dev,use_accelerate=True,dtype=auto",
|
16 |
+
"num_fewshot": [
|
17 |
+
5
|
18 |
+
],
|
19 |
+
"batch_size": 2,
|
20 |
+
"device": "cuda",
|
21 |
+
"no_cache": false,
|
22 |
+
"limit": null,
|
23 |
+
"bootstrap_iters": 100000,
|
24 |
+
"description_dict": null
|
25 |
+
}
|
26 |
+
}
|
benchmark_scores/result_jcommonsenseqa-1.1-0.2.1.json
ADDED
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"jcommonsenseqa-1.1-0.2.1": {
|
4 |
+
"acc": 0.9115281501340483,
|
5 |
+
"acc_stderr": 0.008493108942289001,
|
6 |
+
"acc_norm": 0.837354781054513,
|
7 |
+
"acc_norm_stderr": 0.011037087895481251
|
8 |
+
}
|
9 |
+
},
|
10 |
+
"versions": {
|
11 |
+
"jcommonsenseqa-1.1-0.2.1": 1.1
|
12 |
+
},
|
13 |
+
"config": {
|
14 |
+
"model": "hf-causal-experimental",
|
15 |
+
"model_args": "pretrained=HachiML/youri-2x7b_dev,tokenizer=HachiML/youri-2x7b_dev,use_accelerate=True,dtype=auto",
|
16 |
+
"num_fewshot": [
|
17 |
+
3
|
18 |
+
],
|
19 |
+
"batch_size": 1,
|
20 |
+
"device": "cuda",
|
21 |
+
"no_cache": false,
|
22 |
+
"limit": null,
|
23 |
+
"bootstrap_iters": 100000,
|
24 |
+
"description_dict": null
|
25 |
+
}
|
26 |
+
}
|
benchmark_scores/result_jnli-1.3-0.2.json
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"jnli-1.3-0.2": {
|
4 |
+
"acc": 0.5866885784716516,
|
5 |
+
"acc_stderr": 0.00998323894065562,
|
6 |
+
"acc_norm": 0.5866885784716516,
|
7 |
+
"acc_norm_stderr": 0.00998323894065562,
|
8 |
+
"balanced_acc": 0.7103671626849409,
|
9 |
+
"mcc": 0.48037594993558813,
|
10 |
+
"mcc_stderr": 0.01202764597359472,
|
11 |
+
"macro_f1": 0.6010291975400543
|
12 |
+
}
|
13 |
+
},
|
14 |
+
"versions": {
|
15 |
+
"jnli-1.3-0.2": 1.3
|
16 |
+
},
|
17 |
+
"config": {
|
18 |
+
"model": "hf-causal-experimental",
|
19 |
+
"model_args": "pretrained=HachiML/youri-2x7b_dev,tokenizer=HachiML/youri-2x7b_dev,use_accelerate=True,dtype=auto",
|
20 |
+
"num_fewshot": [
|
21 |
+
3
|
22 |
+
],
|
23 |
+
"batch_size": 8,
|
24 |
+
"device": "cuda",
|
25 |
+
"no_cache": false,
|
26 |
+
"limit": null,
|
27 |
+
"bootstrap_iters": 100000,
|
28 |
+
"description_dict": null
|
29 |
+
}
|
30 |
+
}
|
benchmark_scores/result_jsquad-1.2-0.2.json
ADDED
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"jsquad-1.2-0.2": {
|
4 |
+
"exact_match": 80.07654209815398,
|
5 |
+
"f1": 91.30275393110973
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"jsquad-1.2-0.2": 1.2
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "hf-causal-experimental",
|
13 |
+
"model_args": "pretrained=HachiML/youri-2x7b_dev,tokenizer=HachiML/youri-2x7b_dev,use_accelerate=True,dtype=auto",
|
14 |
+
"num_fewshot": [
|
15 |
+
2
|
16 |
+
],
|
17 |
+
"batch_size": 1,
|
18 |
+
"device": "cuda",
|
19 |
+
"no_cache": false,
|
20 |
+
"limit": null,
|
21 |
+
"bootstrap_iters": 100000,
|
22 |
+
"description_dict": null
|
23 |
+
}
|
24 |
+
}
|
benchmark_scores/result_marc_ja-1.1-0.2.json
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"marc_ja-1.1-0.2": {
|
4 |
+
"acc": 0.9755925008843297,
|
5 |
+
"acc_stderr": 0.0020523733894530468,
|
6 |
+
"acc_norm": 0.9755925008843297,
|
7 |
+
"acc_norm_stderr": 0.0020523733894530468,
|
8 |
+
"balanced_acc": 0.958966027376291,
|
9 |
+
"mcc": 0.9035669838271105,
|
10 |
+
"mcc_stderr": 0.008066342268937397,
|
11 |
+
"macro_f1": 0.9516669276635559
|
12 |
+
}
|
13 |
+
},
|
14 |
+
"versions": {
|
15 |
+
"marc_ja-1.1-0.2": 1.1
|
16 |
+
},
|
17 |
+
"config": {
|
18 |
+
"model": "hf-causal-experimental",
|
19 |
+
"model_args": "pretrained=HachiML/youri-2x7b_dev,tokenizer=HachiML/youri-2x7b_dev,use_accelerate=True,dtype=auto",
|
20 |
+
"num_fewshot": [
|
21 |
+
0
|
22 |
+
],
|
23 |
+
"batch_size": 2,
|
24 |
+
"device": "cuda",
|
25 |
+
"no_cache": false,
|
26 |
+
"limit": null,
|
27 |
+
"bootstrap_iters": 100000,
|
28 |
+
"description_dict": null
|
29 |
+
}
|
30 |
+
}
|
benchmark_scores/result_mgsm.json
ADDED
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"mgsm": {
|
4 |
+
"acc": 0.248,
|
5 |
+
"acc_stderr": 0.027367497504863555
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"mgsm": 1.0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "hf-causal-experimental",
|
13 |
+
"model_args": "pretrained=HachiML/youri-2x7b_dev,tokenizer=HachiML/youri-2x7b_dev,use_accelerate=True,dtype=auto",
|
14 |
+
"num_fewshot": [
|
15 |
+
5
|
16 |
+
],
|
17 |
+
"batch_size": 2,
|
18 |
+
"device": "cuda",
|
19 |
+
"no_cache": false,
|
20 |
+
"limit": null,
|
21 |
+
"bootstrap_iters": 100000,
|
22 |
+
"description_dict": null
|
23 |
+
}
|
24 |
+
}
|
benchmark_scores/result_xlsum_ja.json
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"xlsum_ja": {
|
4 |
+
"rouge2": 25.628917796629892
|
5 |
+
}
|
6 |
+
},
|
7 |
+
"versions": {
|
8 |
+
"xlsum_ja": 1.0
|
9 |
+
},
|
10 |
+
"config": {
|
11 |
+
"model": "hf-causal-experimental",
|
12 |
+
"model_args": "pretrained=HachiML/youri-2x7b_dev,tokenizer=HachiML/youri-2x7b_dev,use_accelerate=True,dtype=auto",
|
13 |
+
"num_fewshot": [
|
14 |
+
1
|
15 |
+
],
|
16 |
+
"batch_size": 2,
|
17 |
+
"device": "cuda",
|
18 |
+
"no_cache": false,
|
19 |
+
"limit": null,
|
20 |
+
"bootstrap_iters": 100000,
|
21 |
+
"description_dict": null
|
22 |
+
}
|
23 |
+
}
|
benchmark_scores/result_xwinograd_ja.json
ADDED
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"xwinograd_ja": {
|
4 |
+
"acc": 0.8143899895724713,
|
5 |
+
"acc_stderr": 0.012561287517973916
|
6 |
+
}
|
7 |
+
},
|
8 |
+
"versions": {
|
9 |
+
"xwinograd_ja": 1.0
|
10 |
+
},
|
11 |
+
"config": {
|
12 |
+
"model": "hf-causal-experimental",
|
13 |
+
"model_args": "pretrained=HachiML/youri-2x7b_dev,tokenizer=HachiML/youri-2x7b_dev,use_accelerate=True,dtype=auto",
|
14 |
+
"num_fewshot": [
|
15 |
+
0
|
16 |
+
],
|
17 |
+
"batch_size": 2,
|
18 |
+
"device": "cuda",
|
19 |
+
"no_cache": false,
|
20 |
+
"limit": null,
|
21 |
+
"bootstrap_iters": 100000,
|
22 |
+
"description_dict": null
|
23 |
+
}
|
24 |
+
}
|