| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.329, | |
| "acc_stderr": 0.014865395385928354 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.336, | |
| "acc_stderr": 0.014944140233795027 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3383333333333333, | |
| "acc_stderr": 0.013664144006618266 | |
| }, | |
| "cb": { | |
| "acc": 0.48214285714285715, | |
| "acc_stderr": 0.06737697508644648, | |
| "f1": 0.3338011695906433 | |
| }, | |
| "copa": { | |
| "acc": 0.79, | |
| "acc_stderr": 0.040936018074033256 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4697271459868552, | |
| "acc_stderr": 0.004980627287147585, | |
| "acc_norm": 0.6141206930890261, | |
| "acc_norm_stderr": 0.004858074013443988 | |
| }, | |
| "rte": { | |
| "acc": 0.5523465703971119, | |
| "acc_stderr": 0.02993107036293953 | |
| }, | |
| "winogrande": { | |
| "acc": 0.574585635359116, | |
| "acc_stderr": 0.013895257666646378 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7156600748262961, | |
| "acc_stderr": 0.010431614128665253 | |
| }, | |
| "boolq": { | |
| "acc": 0.5660550458715596, | |
| "acc_stderr": 0.008668405003744129 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5993265993265994, | |
| "acc_stderr": 0.01005530447425557, | |
| "acc_norm": 0.5576599326599326, | |
| "acc_norm_stderr": 0.01019133444422085 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2781569965870307, | |
| "acc_stderr": 0.013094469919538805, | |
| "acc_norm": 0.30887372013651876, | |
| "acc_norm_stderr": 0.013501770929344003 | |
| }, | |
| "sciq": { | |
| "acc": 0.835, | |
| "acc_stderr": 0.011743632866916145, | |
| "acc_norm": 0.79, | |
| "acc_norm_stderr": 0.01288666233227453 | |
| }, | |
| "piqa": { | |
| "acc": 0.7470076169749728, | |
| "acc_stderr": 0.01014288869886246, | |
| "acc_norm": 0.7519042437431991, | |
| "acc_norm_stderr": 0.010077118315574706 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |