| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.323, | |
| "acc_stderr": 0.014794927843348633 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.364, | |
| "acc_stderr": 0.015222868840522024 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3375, | |
| "acc_stderr": 0.013655897185463655 | |
| }, | |
| "cb": { | |
| "acc": 0.5178571428571429, | |
| "acc_stderr": 0.06737697508644648, | |
| "f1": 0.3525733525733525 | |
| }, | |
| "copa": { | |
| "acc": 0.65, | |
| "acc_stderr": 0.047937248544110196 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.3941445927106154, | |
| "acc_stderr": 0.004876674814874709, | |
| "acc_norm": 0.48645688109938257, | |
| "acc_norm_stderr": 0.004987950663406552 | |
| }, | |
| "rte": { | |
| "acc": 0.51985559566787, | |
| "acc_stderr": 0.030072723167317177 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5138121546961326, | |
| "acc_stderr": 0.014047122916440412 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6424371993586317, | |
| "acc_stderr": 0.011083341168827792 | |
| }, | |
| "boolq": { | |
| "acc": 0.5, | |
| "acc_stderr": 0.008745054976398168 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5071548821548821, | |
| "acc_stderr": 0.01025873302244637, | |
| "acc_norm": 0.4650673400673401, | |
| "acc_norm_stderr": 0.010234713052723658 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.23976109215017063, | |
| "acc_stderr": 0.012476304127453947, | |
| "acc_norm": 0.257679180887372, | |
| "acc_norm_stderr": 0.0127807705627684 | |
| }, | |
| "sciq": { | |
| "acc": 0.735, | |
| "acc_stderr": 0.013963164754809946, | |
| "acc_norm": 0.705, | |
| "acc_norm_stderr": 0.014428554438445517 | |
| }, | |
| "piqa": { | |
| "acc": 0.7023939064200218, | |
| "acc_stderr": 0.010667353792388213, | |
| "acc_norm": 0.7002176278563657, | |
| "acc_norm_stderr": 0.01068968696713809 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |