| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.353, | |
| "acc_stderr": 0.015120172605483699 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.36, | |
| "acc_stderr": 0.015186527932040127 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3641666666666667, | |
| "acc_stderr": 0.013896714966807265 | |
| }, | |
| "cb": { | |
| "acc": 0.5178571428571429, | |
| "acc_stderr": 0.06737697508644647, | |
| "f1": 0.4129474011826953 | |
| }, | |
| "copa": { | |
| "acc": 0.78, | |
| "acc_stderr": 0.041633319989322626 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.466938856801434, | |
| "acc_stderr": 0.004978861409119803, | |
| "acc_norm": 0.6222863971320454, | |
| "acc_norm_stderr": 0.004838246410786256 | |
| }, | |
| "rte": { | |
| "acc": 0.49097472924187724, | |
| "acc_stderr": 0.030091559826331334 | |
| }, | |
| "winogrande": { | |
| "acc": 0.585635359116022, | |
| "acc_stderr": 0.013844846232268563 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7300908605024051, | |
| "acc_stderr": 0.01026541350322146 | |
| }, | |
| "boolq": { | |
| "acc": 0.6220183486238532, | |
| "acc_stderr": 0.008480656964585248 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6233164983164983, | |
| "acc_stderr": 0.00994284807747617, | |
| "acc_norm": 0.6077441077441077, | |
| "acc_norm_stderr": 0.010018744689650043 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3003412969283277, | |
| "acc_stderr": 0.013395909309956999, | |
| "acc_norm": 0.3148464163822526, | |
| "acc_norm_stderr": 0.01357265770308495 | |
| }, | |
| "sciq": { | |
| "acc": 0.904, | |
| "acc_stderr": 0.009320454434783248, | |
| "acc_norm": 0.898, | |
| "acc_norm_stderr": 0.009575368801653897 | |
| }, | |
| "piqa": { | |
| "acc": 0.7524483133841132, | |
| "acc_stderr": 0.010069703966857106, | |
| "acc_norm": 0.7584330794341676, | |
| "acc_norm_stderr": 0.009986718001804453 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |