| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.349, | |
| "acc_stderr": 0.015080663991563098 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.345, | |
| "acc_stderr": 0.015039986742055242 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3466666666666667, | |
| "acc_stderr": 0.013744022550571956 | |
| }, | |
| "cb": { | |
| "acc": 0.48214285714285715, | |
| "acc_stderr": 0.0673769750864465, | |
| "f1": 0.43206548866926225 | |
| }, | |
| "copa": { | |
| "acc": 0.84, | |
| "acc_stderr": 0.03684529491774709 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5317665803624776, | |
| "acc_stderr": 0.004979700695747948, | |
| "acc_norm": 0.7076279625572595, | |
| "acc_norm_stderr": 0.004539227260397019 | |
| }, | |
| "rte": { | |
| "acc": 0.51985559566787, | |
| "acc_stderr": 0.030072723167317177 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6432517758484609, | |
| "acc_stderr": 0.013463393958028726 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7573490112239444, | |
| "acc_stderr": 0.009913300265342056 | |
| }, | |
| "boolq": { | |
| "acc": 0.6590214067278287, | |
| "acc_stderr": 0.00829097981816109 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6839225589225589, | |
| "acc_stderr": 0.009540440071928283, | |
| "acc_norm": 0.6683501683501684, | |
| "acc_norm_stderr": 0.009660733780923948 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3395904436860068, | |
| "acc_stderr": 0.01383903976282016, | |
| "acc_norm": 0.3660409556313993, | |
| "acc_norm_stderr": 0.014077223108470142 | |
| }, | |
| "sciq": { | |
| "acc": 0.935, | |
| "acc_stderr": 0.007799733061832017, | |
| "acc_norm": 0.929, | |
| "acc_norm_stderr": 0.008125578442487916 | |
| }, | |
| "piqa": { | |
| "acc": 0.7905331882480957, | |
| "acc_stderr": 0.009494302979819794, | |
| "acc_norm": 0.7927094668117519, | |
| "acc_norm_stderr": 0.009457844699952372 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |