| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.36, | |
| "acc_stderr": 0.015186527932040119 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.336, | |
| "acc_stderr": 0.014944140233795018 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3425, | |
| "acc_stderr": 0.013704669762934727 | |
| }, | |
| "cb": { | |
| "acc": 0.35714285714285715, | |
| "acc_stderr": 0.0646095738380922, | |
| "f1": 0.2333333333333334 | |
| }, | |
| "copa": { | |
| "acc": 0.76, | |
| "acc_stderr": 0.04292346959909283 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4508066122286397, | |
| "acc_stderr": 0.004965572246803864, | |
| "acc_norm": 0.5993825931089425, | |
| "acc_norm_stderr": 0.004890221012015067 | |
| }, | |
| "rte": { | |
| "acc": 0.5234657039711191, | |
| "acc_stderr": 0.03006330041190266 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5651144435674822, | |
| "acc_stderr": 0.013932814110418015 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7199358631747729, | |
| "acc_stderr": 0.010383764993920484 | |
| }, | |
| "boolq": { | |
| "acc": 0.537308868501529, | |
| "acc_stderr": 0.008720675606388447 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5909090909090909, | |
| "acc_stderr": 0.010088775152615786, | |
| "acc_norm": 0.5778619528619529, | |
| "acc_norm_stderr": 0.010134620524592271 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.26535836177474403, | |
| "acc_stderr": 0.012902554762313962, | |
| "acc_norm": 0.29180887372013653, | |
| "acc_norm_stderr": 0.013284525292403511 | |
| }, | |
| "sciq": { | |
| "acc": 0.883, | |
| "acc_stderr": 0.01016928780271333, | |
| "acc_norm": 0.877, | |
| "acc_norm_stderr": 0.010391293421849876 | |
| }, | |
| "piqa": { | |
| "acc": 0.7562568008705114, | |
| "acc_stderr": 0.010017199471500616, | |
| "acc_norm": 0.7616974972796517, | |
| "acc_norm_stderr": 0.009940334245876219 | |
| } | |
| }, | |
| "versions": { | |
| "anli_r1": 0, | |
| "anli_r2": 0, | |
| "anli_r3": 0, | |
| "cb": 1, | |
| "copa": 0, | |
| "hellaswag": 0, | |
| "rte": 0, | |
| "winogrande": 0, | |
| "storycloze_2016": 0, | |
| "boolq": 1, | |
| "arc_easy": 0, | |
| "arc_challenge": 0, | |
| "sciq": 0, | |
| "piqa": 0 | |
| } | |
| } |