| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.355, | |
| "acc_stderr": 0.01513949154378053 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.359, | |
| "acc_stderr": 0.015177264224798601 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3308333333333333, | |
| "acc_stderr": 0.013588208070708992 | |
| }, | |
| "cb": { | |
| "acc": 0.48214285714285715, | |
| "acc_stderr": 0.0673769750864465, | |
| "f1": 0.3403298350824588 | |
| }, | |
| "copa": { | |
| "acc": 0.7, | |
| "acc_stderr": 0.046056618647183814 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4303923521210914, | |
| "acc_stderr": 0.004941191607317909, | |
| "acc_norm": 0.5592511451902011, | |
| "acc_norm_stderr": 0.004954622308739016 | |
| }, | |
| "rte": { | |
| "acc": 0.5451263537906137, | |
| "acc_stderr": 0.029973636495415252 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5422257300710339, | |
| "acc_stderr": 0.014002284504422442 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.689470871191876, | |
| "acc_stderr": 0.010700112173178448 | |
| }, | |
| "boolq": { | |
| "acc": 0.5856269113149847, | |
| "acc_stderr": 0.00861586377642113 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5648148148148148, | |
| "acc_stderr": 0.010173216430370927, | |
| "acc_norm": 0.5340909090909091, | |
| "acc_norm_stderr": 0.010235908103438688 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2551194539249147, | |
| "acc_stderr": 0.012739038695202098, | |
| "acc_norm": 0.2858361774744027, | |
| "acc_norm_stderr": 0.013203196088537369 | |
| }, | |
| "sciq": { | |
| "acc": 0.862, | |
| "acc_stderr": 0.010912152632504411, | |
| "acc_norm": 0.796, | |
| "acc_norm_stderr": 0.012749374359024391 | |
| }, | |
| "piqa": { | |
| "acc": 0.7328618063112078, | |
| "acc_stderr": 0.010323440492612426, | |
| "acc_norm": 0.73449401523395, | |
| "acc_norm_stderr": 0.010303308653024432 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |