| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.295, | |
| "acc_stderr": 0.014428554438445512 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.312, | |
| "acc_stderr": 0.014658474370509007 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33416666666666667, | |
| "acc_stderr": 0.013622434813136769 | |
| }, | |
| "cb": { | |
| "acc": 0.21428571428571427, | |
| "acc_stderr": 0.055328333517248834, | |
| "f1": 0.1865942028985507 | |
| }, | |
| "copa": { | |
| "acc": 0.83, | |
| "acc_stderr": 0.03775251680686371 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5261900019916351, | |
| "acc_stderr": 0.0049829315659459545, | |
| "acc_norm": 0.702549292969528, | |
| "acc_norm_stderr": 0.004562022467161891 | |
| }, | |
| "rte": { | |
| "acc": 0.5342960288808665, | |
| "acc_stderr": 0.030025579819366426 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6250986582478295, | |
| "acc_stderr": 0.013605544523788 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7536076964190273, | |
| "acc_stderr": 0.009964727533753548 | |
| }, | |
| "boolq": { | |
| "acc": 0.6626911314984709, | |
| "acc_stderr": 0.008269171495741622 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6805555555555556, | |
| "acc_stderr": 0.009567482017268095, | |
| "acc_norm": 0.6565656565656566, | |
| "acc_norm_stderr": 0.00974381736896003 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3250853242320819, | |
| "acc_stderr": 0.013688147309729122, | |
| "acc_norm": 0.34812286689419797, | |
| "acc_norm_stderr": 0.013921008595179333 | |
| }, | |
| "sciq": { | |
| "acc": 0.933, | |
| "acc_stderr": 0.007910345983177549, | |
| "acc_norm": 0.92, | |
| "acc_norm_stderr": 0.008583336977753655 | |
| }, | |
| "piqa": { | |
| "acc": 0.7709466811751904, | |
| "acc_stderr": 0.009804509865175504, | |
| "acc_norm": 0.7856365614798694, | |
| "acc_norm_stderr": 0.009574842136050964 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |