| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.333, | |
| "acc_stderr": 0.014910846164229873 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.351, | |
| "acc_stderr": 0.015100563798316403 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.01363087184382148 | |
| }, | |
| "cb": { | |
| "acc": 0.39285714285714285, | |
| "acc_stderr": 0.0658538889806635, | |
| "f1": 0.27474323062558353 | |
| }, | |
| "copa": { | |
| "acc": 0.59, | |
| "acc_stderr": 0.04943110704237102 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.28281218880701053, | |
| "acc_stderr": 0.004494454911844635, | |
| "acc_norm": 0.2951603266281617, | |
| "acc_norm_stderr": 0.004551826272978058 | |
| }, | |
| "rte": { | |
| "acc": 0.48375451263537905, | |
| "acc_stderr": 0.030080573208738064 | |
| }, | |
| "winogrande": { | |
| "acc": 0.49013417521704816, | |
| "acc_stderr": 0.014049749833367592 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.564404061998931, | |
| "acc_stderr": 0.011466111817562836 | |
| }, | |
| "boolq": { | |
| "acc": 0.4938837920489297, | |
| "acc_stderr": 0.00874440068189348 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.3918350168350168, | |
| "acc_stderr": 0.010016835016834962, | |
| "acc_norm": 0.3653198653198653, | |
| "acc_norm_stderr": 0.009880576614806928 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.18771331058020477, | |
| "acc_stderr": 0.01141100131415512, | |
| "acc_norm": 0.2158703071672355, | |
| "acc_norm_stderr": 0.01202297536003066 | |
| }, | |
| "sciq": { | |
| "acc": 0.692, | |
| "acc_stderr": 0.014606483127342761, | |
| "acc_norm": 0.64, | |
| "acc_norm_stderr": 0.01518652793204012 | |
| }, | |
| "piqa": { | |
| "acc": 0.6381936887921654, | |
| "acc_stderr": 0.011211397313020366, | |
| "acc_norm": 0.6207834602829162, | |
| "acc_norm_stderr": 0.011320331012905074 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |