| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.345, | |
| "acc_stderr": 0.015039986742055235 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.328, | |
| "acc_stderr": 0.014853842487270334 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.36083333333333334, | |
| "acc_stderr": 0.013869180252444862 | |
| }, | |
| "cb": { | |
| "acc": 0.44642857142857145, | |
| "acc_stderr": 0.06703189227942398, | |
| "f1": 0.3346801346801347 | |
| }, | |
| "copa": { | |
| "acc": 0.77, | |
| "acc_stderr": 0.04229525846816506 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.44074885480979886, | |
| "acc_stderr": 0.004954622308739005, | |
| "acc_norm": 0.5762796255725952, | |
| "acc_norm_stderr": 0.004931372657129803 | |
| }, | |
| "rte": { | |
| "acc": 0.48375451263537905, | |
| "acc_stderr": 0.030080573208738064 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5524861878453039, | |
| "acc_stderr": 0.013974847640536204 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6942811330839124, | |
| "acc_stderr": 0.010653884866190597 | |
| }, | |
| "boolq": { | |
| "acc": 0.6061162079510704, | |
| "acc_stderr": 0.008545835792614984 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5963804713804713, | |
| "acc_stderr": 0.01006736896034822, | |
| "acc_norm": 0.5723905723905723, | |
| "acc_norm_stderr": 0.010151683397430673 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2738907849829352, | |
| "acc_stderr": 0.013032004972989505, | |
| "acc_norm": 0.2935153583617747, | |
| "acc_norm_stderr": 0.013307250444941113 | |
| }, | |
| "sciq": { | |
| "acc": 0.879, | |
| "acc_stderr": 0.01031821038094609, | |
| "acc_norm": 0.87, | |
| "acc_norm_stderr": 0.010640169792499361 | |
| }, | |
| "piqa": { | |
| "acc": 0.7285092491838956, | |
| "acc_stderr": 0.010376251176596135, | |
| "acc_norm": 0.7377584330794341, | |
| "acc_norm_stderr": 0.010262502565172442 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |