| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.354, | |
| "acc_stderr": 0.015129868238451772 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.36, | |
| "acc_stderr": 0.015186527932040122 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.365, | |
| "acc_stderr": 0.013903485981413582 | |
| }, | |
| "cb": { | |
| "acc": 0.48214285714285715, | |
| "acc_stderr": 0.0673769750864465, | |
| "f1": 0.33071988595866 | |
| }, | |
| "copa": { | |
| "acc": 0.71, | |
| "acc_stderr": 0.045604802157206845 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4561840270862378, | |
| "acc_stderr": 0.004970585328297622, | |
| "acc_norm": 0.5903206532563234, | |
| "acc_norm_stderr": 0.004907694727935688 | |
| }, | |
| "rte": { | |
| "acc": 0.5054151624548736, | |
| "acc_stderr": 0.03009469812323996 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5406471981057617, | |
| "acc_stderr": 0.014005973823825124 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6990913949759487, | |
| "acc_stderr": 0.010606289538707334 | |
| }, | |
| "boolq": { | |
| "acc": 0.5116207951070336, | |
| "acc_stderr": 0.008742692742551265 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5269360269360269, | |
| "acc_stderr": 0.01024488474062011, | |
| "acc_norm": 0.5122053872053872, | |
| "acc_norm_stderr": 0.010256726235129012 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.27559726962457337, | |
| "acc_stderr": 0.013057169655761838, | |
| "acc_norm": 0.30119453924914674, | |
| "acc_norm_stderr": 0.013406741767847617 | |
| }, | |
| "sciq": { | |
| "acc": 0.817, | |
| "acc_stderr": 0.012233587399477823, | |
| "acc_norm": 0.79, | |
| "acc_norm_stderr": 0.01288666233227455 | |
| }, | |
| "piqa": { | |
| "acc": 0.720348204570185, | |
| "acc_stderr": 0.01047189953030656, | |
| "acc_norm": 0.7252448313384113, | |
| "acc_norm_stderr": 0.010415033676676056 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |