Upload mk7_10_shot_actor.json
Browse files- eval/mk7_10_shot_actor.json +47 -0
eval/mk7_10_shot_actor.json
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"results": {
|
| 3 |
+
"kobest_boolq": {
|
| 4 |
+
"acc": 0.5968660968660968,
|
| 5 |
+
"acc_stderr": 0.013095867778461178,
|
| 6 |
+
"macro_f1": 0.5304603080064664,
|
| 7 |
+
"macro_f1_stderr": 0.013586196223496537
|
| 8 |
+
},
|
| 9 |
+
"kobest_copa": {
|
| 10 |
+
"acc": 0.794,
|
| 11 |
+
"acc_stderr": 0.012795613612786553,
|
| 12 |
+
"macro_f1": 0.7937326775501048,
|
| 13 |
+
"macro_f1_stderr": 0.012808081534088474
|
| 14 |
+
},
|
| 15 |
+
"kobest_hellaswag": {
|
| 16 |
+
"acc": 0.498,
|
| 17 |
+
"acc_stderr": 0.022382894986483524,
|
| 18 |
+
"acc_norm": 0.59,
|
| 19 |
+
"acc_norm_stderr": 0.022017482578127676,
|
| 20 |
+
"macro_f1": 0.493789564405272,
|
| 21 |
+
"macro_f1_stderr": 0.022368543611080722
|
| 22 |
+
},
|
| 23 |
+
"kobest_sentineg": {
|
| 24 |
+
"acc": 0.9193954659949622,
|
| 25 |
+
"acc_stderr": 0.013679912566808524,
|
| 26 |
+
"macro_f1": 0.9191692965489158,
|
| 27 |
+
"macro_f1_stderr": 0.013704064312276852
|
| 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=/home/work/DeepSpeedExamples/DeepSpeedExamples/applications/DeepSpeed-Chat/training/step3_rlhf_finetuning/output/actor",
|
| 39 |
+
"num_fewshot": 10,
|
| 40 |
+
"batch_size": "8",
|
| 41 |
+
"device": "cuda:0",
|
| 42 |
+
"no_cache": true,
|
| 43 |
+
"limit": null,
|
| 44 |
+
"bootstrap_iters": 100000,
|
| 45 |
+
"description_dict": {}
|
| 46 |
+
}
|
| 47 |
+
}
|