| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.329, | |
| "acc_stderr": 0.014865395385928362 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.324, | |
| "acc_stderr": 0.014806864733738854 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3325, | |
| "acc_stderr": 0.013605417345710528 | |
| }, | |
| "cb": { | |
| "acc": 0.4642857142857143, | |
| "acc_stderr": 0.06724777654937658, | |
| "f1": 0.3284421618977745 | |
| }, | |
| "copa": { | |
| "acc": 0.61, | |
| "acc_stderr": 0.04902071300001974 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.27962557259510057, | |
| "acc_stderr": 0.004478979795506768, | |
| "acc_norm": 0.2967536347341167, | |
| "acc_norm_stderr": 0.00455893382299554 | |
| }, | |
| "rte": { | |
| "acc": 0.5379061371841155, | |
| "acc_stderr": 0.030009848912529113 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5019731649565904, | |
| "acc_stderr": 0.014052376259225636 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.5713522180652058, | |
| "acc_stderr": 0.011444094780077097 | |
| }, | |
| "boolq": { | |
| "acc": 0.5293577981651376, | |
| "acc_stderr": 0.008729967580199218 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.39604377104377103, | |
| "acc_stderr": 0.010035580962097935, | |
| "acc_norm": 0.375, | |
| "acc_norm_stderr": 0.009933992677987828 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.18771331058020477, | |
| "acc_stderr": 0.011411001314155117, | |
| "acc_norm": 0.23378839590443687, | |
| "acc_norm_stderr": 0.012368225378507137 | |
| }, | |
| "sciq": { | |
| "acc": 0.681, | |
| "acc_stderr": 0.01474640486547349, | |
| "acc_norm": 0.628, | |
| "acc_norm_stderr": 0.015292149942040577 | |
| }, | |
| "piqa": { | |
| "acc": 0.6240478781284005, | |
| "acc_stderr": 0.011301098166895729, | |
| "acc_norm": 0.6316648531011969, | |
| "acc_norm_stderr": 0.011254089354334357 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |