| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.351, | |
| "acc_stderr": 0.015100563798316405 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.345, | |
| "acc_stderr": 0.015039986742055237 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.345, | |
| "acc_stderr": 0.013728421539454878 | |
| }, | |
| "cb": { | |
| "acc": 0.5714285714285714, | |
| "acc_stderr": 0.06672848092813058, | |
| "f1": 0.37671957671957673 | |
| }, | |
| "copa": { | |
| "acc": 0.78, | |
| "acc_stderr": 0.04163331998932261 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4827723561043617, | |
| "acc_stderr": 0.004986818680313444, | |
| "acc_norm": 0.6446922923720374, | |
| "acc_norm_stderr": 0.004776283203468094 | |
| }, | |
| "rte": { | |
| "acc": 0.5776173285198556, | |
| "acc_stderr": 0.02973162264649588 | |
| }, | |
| "winogrande": { | |
| "acc": 0.595895816890292, | |
| "acc_stderr": 0.013791610664670845 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7252805986103688, | |
| "acc_stderr": 0.010322309878339507 | |
| }, | |
| "boolq": { | |
| "acc": 0.6146788990825688, | |
| "acc_stderr": 0.008511930879680652 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6300505050505051, | |
| "acc_stderr": 0.009906656266021155, | |
| "acc_norm": 0.6111111111111112, | |
| "acc_norm_stderr": 0.01000324833531377 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.30716723549488056, | |
| "acc_stderr": 0.013481034054980945, | |
| "acc_norm": 0.32337883959044367, | |
| "acc_norm_stderr": 0.013669421630012122 | |
| }, | |
| "sciq": { | |
| "acc": 0.919, | |
| "acc_stderr": 0.008632121032139978, | |
| "acc_norm": 0.907, | |
| "acc_norm_stderr": 0.009188875634996669 | |
| }, | |
| "piqa": { | |
| "acc": 0.7529923830250272, | |
| "acc_stderr": 0.010062268140772625, | |
| "acc_norm": 0.7671381936887922, | |
| "acc_norm_stderr": 0.009861236071080753 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |