| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.334, | |
| "acc_stderr": 0.01492201952373296 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.334, | |
| "acc_stderr": 0.014922019523732965 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3566666666666667, | |
| "acc_stderr": 0.013833742805050715 | |
| }, | |
| "cb": { | |
| "acc": 0.2857142857142857, | |
| "acc_stderr": 0.06091449038731725, | |
| "f1": 0.2883181980926342 | |
| }, | |
| "copa": { | |
| "acc": 0.71, | |
| "acc_stderr": 0.04560480215720684 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.43905596494722166, | |
| "acc_stderr": 0.0049525768633152155, | |
| "acc_norm": 0.5635331607249552, | |
| "acc_norm_stderr": 0.004949335356881862 | |
| }, | |
| "rte": { | |
| "acc": 0.5379061371841155, | |
| "acc_stderr": 0.030009848912529117 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5445935280189423, | |
| "acc_stderr": 0.013996485037729793 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.703901656867985, | |
| "acc_stderr": 0.010557307688475126 | |
| }, | |
| "boolq": { | |
| "acc": 0.5758409785932722, | |
| "acc_stderr": 0.00864386902338812 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5387205387205387, | |
| "acc_stderr": 0.010228972678389611, | |
| "acc_norm": 0.48947811447811446, | |
| "acc_norm_stderr": 0.010257511546488232 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.26109215017064846, | |
| "acc_stderr": 0.012835523909473848, | |
| "acc_norm": 0.26706484641638223, | |
| "acc_norm_stderr": 0.012928933196496344 | |
| }, | |
| "sciq": { | |
| "acc": 0.777, | |
| "acc_stderr": 0.013169830843425677, | |
| "acc_norm": 0.686, | |
| "acc_norm_stderr": 0.01468399195108796 | |
| }, | |
| "piqa": { | |
| "acc": 0.7154515778019587, | |
| "acc_stderr": 0.010527218464130614, | |
| "acc_norm": 0.735038084874864, | |
| "acc_norm_stderr": 0.010296557993316038 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |