| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.334, | |
| "acc_stderr": 0.014922019523732958 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.326, | |
| "acc_stderr": 0.014830507204541033 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.315, | |
| "acc_stderr": 0.013415009084004871 | |
| }, | |
| "cb": { | |
| "acc": 0.4107142857142857, | |
| "acc_stderr": 0.06633634150359541, | |
| "f1": 0.28315412186379935 | |
| }, | |
| "copa": { | |
| "acc": 0.85, | |
| "acc_stderr": 0.03588702812826373 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5351523600876319, | |
| "acc_stderr": 0.004977434505403355, | |
| "acc_norm": 0.711611232822147, | |
| "acc_norm_stderr": 0.004520870679457054 | |
| }, | |
| "rte": { | |
| "acc": 0.5523465703971119, | |
| "acc_stderr": 0.02993107036293953 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6361483820047356, | |
| "acc_stderr": 0.013521488896883413 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7584179583110636, | |
| "acc_stderr": 0.009898418790766705 | |
| }, | |
| "boolq": { | |
| "acc": 0.6611620795107034, | |
| "acc_stderr": 0.008278325755273739 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6898148148148148, | |
| "acc_stderr": 0.009491721291998517, | |
| "acc_norm": 0.6679292929292929, | |
| "acc_norm_stderr": 0.009663817543072694 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3438566552901024, | |
| "acc_stderr": 0.013880644570156205, | |
| "acc_norm": 0.3609215017064846, | |
| "acc_norm_stderr": 0.01403476138617546 | |
| }, | |
| "sciq": { | |
| "acc": 0.944, | |
| "acc_stderr": 0.007274401481697059, | |
| "acc_norm": 0.931, | |
| "acc_norm_stderr": 0.008018934050315157 | |
| }, | |
| "piqa": { | |
| "acc": 0.779651795429815, | |
| "acc_stderr": 0.00967053545685313, | |
| "acc_norm": 0.7889009793253536, | |
| "acc_norm_stderr": 0.009521377378734144 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |