| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.357, | |
| "acc_stderr": 0.015158521721486774 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.334, | |
| "acc_stderr": 0.014922019523732967 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.37416666666666665, | |
| "acc_stderr": 0.01397501560175897 | |
| }, | |
| "cb": { | |
| "acc": 0.39285714285714285, | |
| "acc_stderr": 0.0658538889806635, | |
| "f1": 0.26788664379209554 | |
| }, | |
| "copa": { | |
| "acc": 0.75, | |
| "acc_stderr": 0.04351941398892446 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.45339573790081655, | |
| "acc_stderr": 0.004968058944472161, | |
| "acc_norm": 0.5857398924517029, | |
| "acc_norm_stderr": 0.004915870966174404 | |
| }, | |
| "rte": { | |
| "acc": 0.4657039711191336, | |
| "acc_stderr": 0.030025579819366422 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5509076558800315, | |
| "acc_stderr": 0.01397945938914086 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6878674505611972, | |
| "acc_stderr": 0.010715220346279681 | |
| }, | |
| "boolq": { | |
| "acc": 0.517125382262997, | |
| "acc_stderr": 0.008739923994130054 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5307239057239057, | |
| "acc_stderr": 0.010240395584815237, | |
| "acc_norm": 0.5084175084175084, | |
| "acc_norm_stderr": 0.010258329515226462 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2568259385665529, | |
| "acc_stderr": 0.0127669237941168, | |
| "acc_norm": 0.31399317406143346, | |
| "acc_norm_stderr": 0.013562691224726297 | |
| }, | |
| "sciq": { | |
| "acc": 0.799, | |
| "acc_stderr": 0.012679107214617328, | |
| "acc_norm": 0.751, | |
| "acc_norm_stderr": 0.0136816002787023 | |
| }, | |
| "piqa": { | |
| "acc": 0.7399347116430903, | |
| "acc_stderr": 0.010234893249061303, | |
| "acc_norm": 0.7328618063112078, | |
| "acc_norm_stderr": 0.01032344049261243 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |