| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.352, | |
| "acc_stderr": 0.015110404505648668 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.327, | |
| "acc_stderr": 0.014842213153411247 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3458333333333333, | |
| "acc_stderr": 0.013736245342311012 | |
| }, | |
| "cb": { | |
| "acc": 0.44642857142857145, | |
| "acc_stderr": 0.06703189227942397, | |
| "f1": 0.3134878193701723 | |
| }, | |
| "copa": { | |
| "acc": 0.79, | |
| "acc_stderr": 0.040936018074033256 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4457279426409082, | |
| "acc_stderr": 0.004960299952519412, | |
| "acc_norm": 0.5867357100179247, | |
| "acc_norm_stderr": 0.00491413085543178 | |
| }, | |
| "rte": { | |
| "acc": 0.5415162454873647, | |
| "acc_stderr": 0.029992535385373314 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5730071033938438, | |
| "acc_stderr": 0.013901878072575055 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7071084981293426, | |
| "acc_stderr": 0.010523873293246309 | |
| }, | |
| "boolq": { | |
| "acc": 0.5960244648318043, | |
| "acc_stderr": 0.008582268854021401 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6153198653198653, | |
| "acc_stderr": 0.00998317170700901, | |
| "acc_norm": 0.5989057239057239, | |
| "acc_norm_stderr": 0.010057051106534385 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.27474402730375425, | |
| "acc_stderr": 0.013044617212771227, | |
| "acc_norm": 0.30887372013651876, | |
| "acc_norm_stderr": 0.013501770929344003 | |
| }, | |
| "sciq": { | |
| "acc": 0.891, | |
| "acc_stderr": 0.009859828407037191, | |
| "acc_norm": 0.863, | |
| "acc_norm_stderr": 0.010878848714333316 | |
| }, | |
| "piqa": { | |
| "acc": 0.7301414581066377, | |
| "acc_stderr": 0.0103565954218522, | |
| "acc_norm": 0.7312295973884657, | |
| "acc_norm_stderr": 0.01034339294009 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |