| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.319, | |
| "acc_stderr": 0.014746404865473493 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.337, | |
| "acc_stderr": 0.0149550879186536 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3375, | |
| "acc_stderr": 0.013655897185463665 | |
| }, | |
| "cb": { | |
| "acc": 0.44642857142857145, | |
| "acc_stderr": 0.06703189227942397, | |
| "f1": 0.3011063011063011 | |
| }, | |
| "copa": { | |
| "acc": 0.54, | |
| "acc_stderr": 0.05009082659620332 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.27853017327225654, | |
| "acc_stderr": 0.004473595650807679, | |
| "acc_norm": 0.2961561441943836, | |
| "acc_norm_stderr": 0.004556276293751938 | |
| }, | |
| "rte": { | |
| "acc": 0.5054151624548736, | |
| "acc_stderr": 0.030094698123239966 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5067087608524072, | |
| "acc_stderr": 0.014051220692330352 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.5777659005879209, | |
| "acc_stderr": 0.011421727692385657 | |
| }, | |
| "boolq": { | |
| "acc": 0.5434250764525994, | |
| "acc_stderr": 0.008712010793695303 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.39604377104377103, | |
| "acc_stderr": 0.010035580962097937, | |
| "acc_norm": 0.36153198653198654, | |
| "acc_norm_stderr": 0.009858506543162062 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.18515358361774745, | |
| "acc_stderr": 0.011350774438389699, | |
| "acc_norm": 0.22696245733788395, | |
| "acc_norm_stderr": 0.012240491536132872 | |
| }, | |
| "sciq": { | |
| "acc": 0.693, | |
| "acc_stderr": 0.014593284892852623, | |
| "acc_norm": 0.635, | |
| "acc_norm_stderr": 0.015231776226264903 | |
| }, | |
| "piqa": { | |
| "acc": 0.6305767138193689, | |
| "acc_stderr": 0.011260988628572347, | |
| "acc_norm": 0.6311207834602829, | |
| "acc_norm_stderr": 0.011257546676908804 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |