| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.313, | |
| "acc_stderr": 0.014671272822977883 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.325, | |
| "acc_stderr": 0.014818724459095529 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33666666666666667, | |
| "acc_stderr": 0.013647602942406394 | |
| }, | |
| "cb": { | |
| "acc": 0.48214285714285715, | |
| "acc_stderr": 0.0673769750864465, | |
| "f1": 0.3995062282572102 | |
| }, | |
| "copa": { | |
| "acc": 0.74, | |
| "acc_stderr": 0.0440844002276808 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4510057757418841, | |
| "acc_stderr": 0.004965768348628053, | |
| "acc_norm": 0.5970922127066322, | |
| "acc_norm_stderr": 0.004894801119898596 | |
| }, | |
| "rte": { | |
| "acc": 0.5090252707581228, | |
| "acc_stderr": 0.030091559826331334 | |
| }, | |
| "winogrande": { | |
| "acc": 0.569060773480663, | |
| "acc_stderr": 0.013917796623335966 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7124532335649385, | |
| "acc_stderr": 0.010466744473098368 | |
| }, | |
| "boolq": { | |
| "acc": 0.6204892966360857, | |
| "acc_stderr": 0.008487341975756834 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6405723905723906, | |
| "acc_stderr": 0.009845958893373764, | |
| "acc_norm": 0.6376262626262627, | |
| "acc_norm_stderr": 0.009863468202583773 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2858361774744027, | |
| "acc_stderr": 0.013203196088537369, | |
| "acc_norm": 0.32081911262798635, | |
| "acc_norm_stderr": 0.013640943091946524 | |
| }, | |
| "sciq": { | |
| "acc": 0.923, | |
| "acc_stderr": 0.008434580140240651, | |
| "acc_norm": 0.925, | |
| "acc_norm_stderr": 0.00833333333333335 | |
| }, | |
| "piqa": { | |
| "acc": 0.7442872687704026, | |
| "acc_stderr": 0.010178690109459862, | |
| "acc_norm": 0.7519042437431991, | |
| "acc_norm_stderr": 0.010077118315574703 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |