| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.325, | |
| "acc_stderr": 0.014818724459095524 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.337, | |
| "acc_stderr": 0.0149550879186536 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3475, | |
| "acc_stderr": 0.013751753243291852 | |
| }, | |
| "cb": { | |
| "acc": 0.4107142857142857, | |
| "acc_stderr": 0.06633634150359538, | |
| "f1": 0.3114633159610671 | |
| }, | |
| "copa": { | |
| "acc": 0.85, | |
| "acc_stderr": 0.035887028128263714 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5340569607647879, | |
| "acc_stderr": 0.0049781928934062745, | |
| "acc_norm": 0.7103166699860586, | |
| "acc_norm_stderr": 0.00452688302102762 | |
| }, | |
| "rte": { | |
| "acc": 0.5415162454873647, | |
| "acc_stderr": 0.029992535385373314 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6337805840568271, | |
| "acc_stderr": 0.013540144376588896 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7659005879208979, | |
| "acc_stderr": 0.009791868211495318 | |
| }, | |
| "boolq": { | |
| "acc": 0.6321100917431193, | |
| "acc_stderr": 0.008434276591093038 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6893939393939394, | |
| "acc_stderr": 0.009495260551195607, | |
| "acc_norm": 0.6721380471380471, | |
| "acc_norm_stderr": 0.00963258707617002 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.34982935153583616, | |
| "acc_stderr": 0.01393680921215828, | |
| "acc_norm": 0.3626279863481229, | |
| "acc_norm_stderr": 0.014049106564955012 | |
| }, | |
| "sciq": { | |
| "acc": 0.934, | |
| "acc_stderr": 0.007855297938697589, | |
| "acc_norm": 0.933, | |
| "acc_norm_stderr": 0.007910345983177549 | |
| }, | |
| "piqa": { | |
| "acc": 0.7780195865070729, | |
| "acc_stderr": 0.009696120744662026, | |
| "acc_norm": 0.7878128400435256, | |
| "acc_norm_stderr": 0.009539299828174055 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |