| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.343, | |
| "acc_stderr": 0.015019206922356951 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.346, | |
| "acc_stderr": 0.01505026612756445 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.36083333333333334, | |
| "acc_stderr": 0.01386918025244486 | |
| }, | |
| "cb": { | |
| "acc": 0.5535714285714286, | |
| "acc_stderr": 0.06703189227942395, | |
| "f1": 0.4583333333333333 | |
| }, | |
| "copa": { | |
| "acc": 0.8, | |
| "acc_stderr": 0.040201512610368445 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.48157737502489545, | |
| "acc_stderr": 0.0049863932662691625, | |
| "acc_norm": 0.6417048396733719, | |
| "acc_norm_stderr": 0.00478519504988916 | |
| }, | |
| "rte": { | |
| "acc": 0.5379061371841155, | |
| "acc_stderr": 0.030009848912529113 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6085240726124704, | |
| "acc_stderr": 0.01371748707129085 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7338321753073223, | |
| "acc_stderr": 0.010220104800551206 | |
| }, | |
| "boolq": { | |
| "acc": 0.6119266055045871, | |
| "acc_stderr": 0.00852313058476084 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6283670033670034, | |
| "acc_stderr": 0.00991589712365879, | |
| "acc_norm": 0.6153198653198653, | |
| "acc_norm_stderr": 0.009983171707008997 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2960750853242321, | |
| "acc_stderr": 0.013340916085246271, | |
| "acc_norm": 0.3242320819112628, | |
| "acc_norm_stderr": 0.013678810399518819 | |
| }, | |
| "sciq": { | |
| "acc": 0.923, | |
| "acc_stderr": 0.008434580140240648, | |
| "acc_norm": 0.912, | |
| "acc_norm_stderr": 0.008963053962592074 | |
| }, | |
| "piqa": { | |
| "acc": 0.7595212187159956, | |
| "acc_stderr": 0.009971345364651078, | |
| "acc_norm": 0.7676822633297062, | |
| "acc_norm_stderr": 0.009853201384168243 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |