| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.014933117490932577 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.337, | |
| "acc_stderr": 0.014955087918653609 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3325, | |
| "acc_stderr": 0.013605417345710528 | |
| }, | |
| "cb": { | |
| "acc": 0.23214285714285715, | |
| "acc_stderr": 0.056929390240001085, | |
| "f1": 0.23148148148148148 | |
| }, | |
| "copa": { | |
| "acc": 0.84, | |
| "acc_stderr": 0.0368452949177471 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.47470623381796456, | |
| "acc_stderr": 0.004983392650570959, | |
| "acc_norm": 0.6319458275243975, | |
| "acc_norm_stderr": 0.004812905279066442 | |
| }, | |
| "rte": { | |
| "acc": 0.47653429602888087, | |
| "acc_stderr": 0.030063300411902652 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6108918705603789, | |
| "acc_stderr": 0.013702520871485949 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.721004810261892, | |
| "acc_stderr": 0.010371620932652795 | |
| }, | |
| "boolq": { | |
| "acc": 0.6192660550458715, | |
| "acc_stderr": 0.008492625561656215 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6380471380471381, | |
| "acc_stderr": 0.00986099146668847, | |
| "acc_norm": 0.6216329966329966, | |
| "acc_norm_stderr": 0.00995157568333195 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.29948805460750855, | |
| "acc_stderr": 0.013385021637313563, | |
| "acc_norm": 0.3361774744027304, | |
| "acc_norm_stderr": 0.013804855026205763 | |
| }, | |
| "sciq": { | |
| "acc": 0.915, | |
| "acc_stderr": 0.008823426366942323, | |
| "acc_norm": 0.91, | |
| "acc_norm_stderr": 0.009054390204866444 | |
| }, | |
| "piqa": { | |
| "acc": 0.7529923830250272, | |
| "acc_stderr": 0.010062268140772625, | |
| "acc_norm": 0.7568008705114254, | |
| "acc_norm_stderr": 0.010009611953858917 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |