| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.349, | |
| "acc_stderr": 0.0150806639915631 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.014933117490932566 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.34833333333333333, | |
| "acc_stderr": 0.013759437498874075 | |
| }, | |
| "cb": { | |
| "acc": 0.375, | |
| "acc_stderr": 0.06527912098338669, | |
| "f1": 0.25396825396825395 | |
| }, | |
| "copa": { | |
| "acc": 0.74, | |
| "acc_stderr": 0.044084400227680794 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.45797649870543716, | |
| "acc_stderr": 0.004972126523031945, | |
| "acc_norm": 0.5928101971718781, | |
| "acc_norm_stderr": 0.004903066639761946 | |
| }, | |
| "rte": { | |
| "acc": 0.4981949458483754, | |
| "acc_stderr": 0.030096267148976626 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5651144435674822, | |
| "acc_stderr": 0.013932814110418013 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6857295563869589, | |
| "acc_stderr": 0.010735132285108176 | |
| }, | |
| "boolq": { | |
| "acc": 0.5137614678899083, | |
| "acc_stderr": 0.008741742106878655 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5294612794612794, | |
| "acc_stderr": 0.010241957728409683, | |
| "acc_norm": 0.5105218855218855, | |
| "acc_norm_stderr": 0.010257511546488227 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.26791808873720135, | |
| "acc_stderr": 0.012942030195136421, | |
| "acc_norm": 0.3054607508532423, | |
| "acc_norm_stderr": 0.013460080478002501 | |
| }, | |
| "sciq": { | |
| "acc": 0.799, | |
| "acc_stderr": 0.012679107214617331, | |
| "acc_norm": 0.778, | |
| "acc_norm_stderr": 0.013148721948877366 | |
| }, | |
| "piqa": { | |
| "acc": 0.7328618063112078, | |
| "acc_stderr": 0.010323440492612437, | |
| "acc_norm": 0.7377584330794341, | |
| "acc_norm_stderr": 0.010262502565172447 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |