| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.367, | |
| "acc_stderr": 0.01524937846417175 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.325, | |
| "acc_stderr": 0.014818724459095527 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3425, | |
| "acc_stderr": 0.013704669762934722 | |
| }, | |
| "cb": { | |
| "acc": 0.42857142857142855, | |
| "acc_stderr": 0.06672848092813058, | |
| "f1": 0.3487396784006953 | |
| }, | |
| "copa": { | |
| "acc": 0.71, | |
| "acc_stderr": 0.04560480215720684 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.46265684126667994, | |
| "acc_stderr": 0.004975845335086618, | |
| "acc_norm": 0.6078470424218283, | |
| "acc_norm_stderr": 0.004872326888655505 | |
| }, | |
| "rte": { | |
| "acc": 0.5379061371841155, | |
| "acc_stderr": 0.030009848912529117 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5572217837411207, | |
| "acc_stderr": 0.013960157350784985 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7113842864778194, | |
| "acc_stderr": 0.010478311785642947 | |
| }, | |
| "boolq": { | |
| "acc": 0.6192660550458715, | |
| "acc_stderr": 0.008492625561656217 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6056397306397306, | |
| "acc_stderr": 0.010028176038393007, | |
| "acc_norm": 0.5812289562289562, | |
| "acc_norm_stderr": 0.010123487160167819 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.28498293515358364, | |
| "acc_stderr": 0.013191348179838795, | |
| "acc_norm": 0.3191126279863481, | |
| "acc_norm_stderr": 0.013621696119173304 | |
| }, | |
| "sciq": { | |
| "acc": 0.865, | |
| "acc_stderr": 0.010811655372416051, | |
| "acc_norm": 0.851, | |
| "acc_norm_stderr": 0.01126614068463217 | |
| }, | |
| "piqa": { | |
| "acc": 0.7453754080522307, | |
| "acc_stderr": 0.01016443223706048, | |
| "acc_norm": 0.7453754080522307, | |
| "acc_norm_stderr": 0.010164432237060492 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |