| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.311, | |
| "acc_stderr": 0.014645596385722695 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.313, | |
| "acc_stderr": 0.014671272822977886 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33166666666666667, | |
| "acc_stderr": 0.013596836729485156 | |
| }, | |
| "cb": { | |
| "acc": 0.3392857142857143, | |
| "acc_stderr": 0.06384226561930828, | |
| "f1": 0.23827865281885505 | |
| }, | |
| "copa": { | |
| "acc": 0.8, | |
| "acc_stderr": 0.04020151261036845 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.46036646086436966, | |
| "acc_stderr": 0.0049740806383642665, | |
| "acc_norm": 0.6048595897231627, | |
| "acc_norm_stderr": 0.00487881696101204 | |
| }, | |
| "rte": { | |
| "acc": 0.4981949458483754, | |
| "acc_stderr": 0.030096267148976633 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5580110497237569, | |
| "acc_stderr": 0.013957584079109001 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6755745590593266, | |
| "acc_stderr": 0.01082613134499089 | |
| }, | |
| "boolq": { | |
| "acc": 0.6061162079510704, | |
| "acc_stderr": 0.008545835792614982 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5555555555555556, | |
| "acc_stderr": 0.01019625483869168, | |
| "acc_norm": 0.5366161616161617, | |
| "acc_norm_stderr": 0.01023223506393303 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2568259385665529, | |
| "acc_stderr": 0.0127669237941168, | |
| "acc_norm": 0.30119453924914674, | |
| "acc_norm_stderr": 0.01340674176784762 | |
| }, | |
| "sciq": { | |
| "acc": 0.852, | |
| "acc_stderr": 0.011234866364235239, | |
| "acc_norm": 0.834, | |
| "acc_norm_stderr": 0.011772110370812185 | |
| }, | |
| "piqa": { | |
| "acc": 0.719804134929271, | |
| "acc_stderr": 0.010478122015577082, | |
| "acc_norm": 0.7181719260065288, | |
| "acc_norm_stderr": 0.010496675231258159 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |