| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.31, | |
| "acc_stderr": 0.014632638658632896 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.354, | |
| "acc_stderr": 0.015129868238451773 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3375, | |
| "acc_stderr": 0.013655897185463655 | |
| }, | |
| "cb": { | |
| "acc": 0.44642857142857145, | |
| "acc_stderr": 0.06703189227942398, | |
| "f1": 0.43517730496453905 | |
| }, | |
| "copa": { | |
| "acc": 0.79, | |
| "acc_stderr": 0.040936018074033256 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4761999601672974, | |
| "acc_stderr": 0.004984125363319067, | |
| "acc_norm": 0.6349332802230632, | |
| "acc_norm_stderr": 0.004804649197163698 | |
| }, | |
| "rte": { | |
| "acc": 0.5306859205776173, | |
| "acc_stderr": 0.030039730592197812 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5872138910812944, | |
| "acc_stderr": 0.013837060648682094 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7183324425440941, | |
| "acc_stderr": 0.010401844358587662 | |
| }, | |
| "boolq": { | |
| "acc": 0.6238532110091743, | |
| "acc_stderr": 0.008472516562330725 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.627104377104377, | |
| "acc_stderr": 0.009922743197129253, | |
| "acc_norm": 0.6081649831649831, | |
| "acc_norm_stderr": 0.010016835016834974 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.30631399317406144, | |
| "acc_stderr": 0.013470584417276513, | |
| "acc_norm": 0.32593856655290104, | |
| "acc_norm_stderr": 0.013697432466693246 | |
| }, | |
| "sciq": { | |
| "acc": 0.909, | |
| "acc_stderr": 0.009099549538400245, | |
| "acc_norm": 0.898, | |
| "acc_norm_stderr": 0.009575368801653886 | |
| }, | |
| "piqa": { | |
| "acc": 0.7568008705114254, | |
| "acc_stderr": 0.010009611953858917, | |
| "acc_norm": 0.7665941240478781, | |
| "acc_norm_stderr": 0.009869247889520994 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |