Upload 3 files
Browse files- eval/chatskku_0_shot.json +47 -0
- eval/koalpaca_0_shot.json +47 -0
- eval/kullm_0_shot.json +47 -0
eval/chatskku_0_shot.json
ADDED
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"kobest_boolq": {
|
4 |
+
"acc": 0.5256410256410257,
|
5 |
+
"acc_stderr": 0.013331203323083502,
|
6 |
+
"macro_f1": 0.3992444490131579,
|
7 |
+
"macro_f1_stderr": 0.010971221905948464
|
8 |
+
},
|
9 |
+
"kobest_copa": {
|
10 |
+
"acc": 0.785,
|
11 |
+
"acc_stderr": 0.012997843819031832,
|
12 |
+
"macro_f1": 0.7847052615029975,
|
13 |
+
"macro_f1_stderr": 0.012967330755596378
|
14 |
+
},
|
15 |
+
"kobest_hellaswag": {
|
16 |
+
"acc": 0.466,
|
17 |
+
"acc_stderr": 0.02233126442325838,
|
18 |
+
"acc_norm": 0.592,
|
19 |
+
"acc_norm_stderr": 0.022000910893877193,
|
20 |
+
"macro_f1": 0.46440112486053614,
|
21 |
+
"macro_f1_stderr": 0.022343780241897912
|
22 |
+
},
|
23 |
+
"kobest_sentineg": {
|
24 |
+
"acc": 0.5062972292191436,
|
25 |
+
"acc_stderr": 0.02512395255890725,
|
26 |
+
"macro_f1": 0.34501683501683506,
|
27 |
+
"macro_f1_stderr": 0.01352473866949455
|
28 |
+
}
|
29 |
+
},
|
30 |
+
"versions": {
|
31 |
+
"kobest_boolq": 0,
|
32 |
+
"kobest_copa": 0,
|
33 |
+
"kobest_hellaswag": 0,
|
34 |
+
"kobest_sentineg": 0
|
35 |
+
},
|
36 |
+
"config": {
|
37 |
+
"model": "gpt2",
|
38 |
+
"model_args": "pretrained=jojo0217/ChatSKKU5.8B",
|
39 |
+
"num_fewshot": 0,
|
40 |
+
"batch_size": "4",
|
41 |
+
"device": "cuda:0",
|
42 |
+
"no_cache": true,
|
43 |
+
"limit": null,
|
44 |
+
"bootstrap_iters": 100000,
|
45 |
+
"description_dict": {}
|
46 |
+
}
|
47 |
+
}
|
eval/koalpaca_0_shot.json
ADDED
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"kobest_boolq": {
|
4 |
+
"acc": 0.5626780626780626,
|
5 |
+
"acc_stderr": 0.013243469872682661,
|
6 |
+
"macro_f1": 0.5053421868652208,
|
7 |
+
"macro_f1_stderr": 0.013395027664000827
|
8 |
+
},
|
9 |
+
"kobest_copa": {
|
10 |
+
"acc": 0.698,
|
11 |
+
"acc_stderr": 0.014526080235459544,
|
12 |
+
"macro_f1": 0.6977279551596436,
|
13 |
+
"macro_f1_stderr": 0.014458551946793323
|
14 |
+
},
|
15 |
+
"kobest_hellaswag": {
|
16 |
+
"acc": 0.384,
|
17 |
+
"acc_stderr": 0.021772369465547194,
|
18 |
+
"acc_norm": 0.462,
|
19 |
+
"acc_norm_stderr": 0.022318338119870537,
|
20 |
+
"macro_f1": 0.3834926136948366,
|
21 |
+
"macro_f1_stderr": 0.02180403516376896
|
22 |
+
},
|
23 |
+
"kobest_sentineg": {
|
24 |
+
"acc": 0.3879093198992443,
|
25 |
+
"acc_stderr": 0.024486425812624234,
|
26 |
+
"macro_f1": 0.38484935437589674,
|
27 |
+
"macro_f1_stderr": 0.024320271236487046
|
28 |
+
}
|
29 |
+
},
|
30 |
+
"versions": {
|
31 |
+
"kobest_boolq": 0,
|
32 |
+
"kobest_copa": 0,
|
33 |
+
"kobest_hellaswag": 0,
|
34 |
+
"kobest_sentineg": 0
|
35 |
+
},
|
36 |
+
"config": {
|
37 |
+
"model": "gpt2",
|
38 |
+
"model_args": "pretrained=beomi/KoAlpaca-Polyglot-5.8B",
|
39 |
+
"num_fewshot": 0,
|
40 |
+
"batch_size": "4",
|
41 |
+
"device": "cuda:0",
|
42 |
+
"no_cache": true,
|
43 |
+
"limit": null,
|
44 |
+
"bootstrap_iters": 100000,
|
45 |
+
"description_dict": {}
|
46 |
+
}
|
47 |
+
}
|
eval/kullm_0_shot.json
ADDED
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"results": {
|
3 |
+
"kobest_boolq": {
|
4 |
+
"acc": 0.5028490028490028,
|
5 |
+
"acc_stderr": 0.013348550797680814,
|
6 |
+
"macro_f1": 0.3371320037986705,
|
7 |
+
"macro_f1_stderr": 0.006306119104194062
|
8 |
+
},
|
9 |
+
"kobest_copa": {
|
10 |
+
"acc": 0.766,
|
11 |
+
"acc_stderr": 0.013394902889660007,
|
12 |
+
"macro_f1": 0.7652638674884438,
|
13 |
+
"macro_f1_stderr": 0.01339434219317699
|
14 |
+
},
|
15 |
+
"kobest_hellaswag": {
|
16 |
+
"acc": 0.444,
|
17 |
+
"acc_stderr": 0.02224224437573102,
|
18 |
+
"acc_norm": 0.552,
|
19 |
+
"acc_norm_stderr": 0.02226169729227013,
|
20 |
+
"macro_f1": 0.4425279610634302,
|
21 |
+
"macro_f1_stderr": 0.02226671964725493
|
22 |
+
},
|
23 |
+
"kobest_sentineg": {
|
24 |
+
"acc": 0.5037783375314862,
|
25 |
+
"acc_stderr": 0.025125227983562776,
|
26 |
+
"macro_f1": 0.33500837520938026,
|
27 |
+
"macro_f1_stderr": 0.011139659523368081
|
28 |
+
}
|
29 |
+
},
|
30 |
+
"versions": {
|
31 |
+
"kobest_boolq": 0,
|
32 |
+
"kobest_copa": 0,
|
33 |
+
"kobest_hellaswag": 0,
|
34 |
+
"kobest_sentineg": 0
|
35 |
+
},
|
36 |
+
"config": {
|
37 |
+
"model": "gpt2",
|
38 |
+
"model_args": "pretrained=nlpai-lab/kullm-polyglot-5.8b-v2",
|
39 |
+
"num_fewshot": 0,
|
40 |
+
"batch_size": "2",
|
41 |
+
"device": "cuda:0",
|
42 |
+
"no_cache": true,
|
43 |
+
"limit": null,
|
44 |
+
"bootstrap_iters": 100000,
|
45 |
+
"description_dict": {}
|
46 |
+
}
|
47 |
+
}
|