| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.337, | |
| "acc_stderr": 0.014955087918653591 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.347, | |
| "acc_stderr": 0.015060472031706618 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3233333333333333, | |
| "acc_stderr": 0.01350837286730022 | |
| }, | |
| "cb": { | |
| "acc": 0.375, | |
| "acc_stderr": 0.06527912098338669, | |
| "f1": 0.2554143126177024 | |
| }, | |
| "copa": { | |
| "acc": 0.63, | |
| "acc_stderr": 0.04852365870939099 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.279326827325234, | |
| "acc_stderr": 0.004477514681328155, | |
| "acc_norm": 0.2956582354112727, | |
| "acc_norm_stderr": 0.0045540545376920125 | |
| }, | |
| "rte": { | |
| "acc": 0.5379061371841155, | |
| "acc_stderr": 0.030009848912529117 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5114443567482242, | |
| "acc_stderr": 0.014048804199859325 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.5750935328701229, | |
| "acc_stderr": 0.011431286492205844 | |
| }, | |
| "boolq": { | |
| "acc": 0.5614678899082569, | |
| "acc_stderr": 0.008678720482001873 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.3952020202020202, | |
| "acc_stderr": 0.010031894052790978, | |
| "acc_norm": 0.36826599326599324, | |
| "acc_norm_stderr": 0.009897286209010894 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.1825938566552901, | |
| "acc_stderr": 0.011289730684565, | |
| "acc_norm": 0.2295221843003413, | |
| "acc_norm_stderr": 0.012288926760890797 | |
| }, | |
| "sciq": { | |
| "acc": 0.686, | |
| "acc_stderr": 0.014683991951087955, | |
| "acc_norm": 0.635, | |
| "acc_norm_stderr": 0.015231776226264903 | |
| }, | |
| "piqa": { | |
| "acc": 0.6294885745375408, | |
| "acc_stderr": 0.011267826475447665, | |
| "acc_norm": 0.6322089227421109, | |
| "acc_norm_stderr": 0.01125061664667879 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |