| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.311, | |
| "acc_stderr": 0.014645596385722694 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.299, | |
| "acc_stderr": 0.014484778521220477 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.013630871843821474 | |
| }, | |
| "cb": { | |
| "acc": 0.5178571428571429, | |
| "acc_stderr": 0.06737697508644647, | |
| "f1": 0.43401043401043404 | |
| }, | |
| "copa": { | |
| "acc": 0.82, | |
| "acc_stderr": 0.038612291966536955 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.45140410276837284, | |
| "acc_stderr": 0.004966158142645416, | |
| "acc_norm": 0.601274646484764, | |
| "acc_norm_stderr": 0.0048863535635718415 | |
| }, | |
| "rte": { | |
| "acc": 0.49097472924187724, | |
| "acc_stderr": 0.030091559826331334 | |
| }, | |
| "winogrande": { | |
| "acc": 0.590370955011839, | |
| "acc_stderr": 0.013821049109655491 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7097808658471406, | |
| "acc_stderr": 0.010495529690730063 | |
| }, | |
| "boolq": { | |
| "acc": 0.6241590214067279, | |
| "acc_stderr": 0.008471147248160114 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6401515151515151, | |
| "acc_stderr": 0.009848484848484843, | |
| "acc_norm": 0.6346801346801347, | |
| "acc_norm_stderr": 0.009880576614806924 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.28924914675767915, | |
| "acc_stderr": 0.013250012579393443, | |
| "acc_norm": 0.318259385665529, | |
| "acc_norm_stderr": 0.013611993916971453 | |
| }, | |
| "sciq": { | |
| "acc": 0.927, | |
| "acc_stderr": 0.008230354715244055, | |
| "acc_norm": 0.928, | |
| "acc_norm_stderr": 0.008178195576218681 | |
| }, | |
| "piqa": { | |
| "acc": 0.7453754080522307, | |
| "acc_stderr": 0.010164432237060487, | |
| "acc_norm": 0.7448313384113167, | |
| "acc_norm_stderr": 0.010171571592521834 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |