| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.329, | |
| "acc_stderr": 0.014865395385928367 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.312, | |
| "acc_stderr": 0.014658474370509005 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3458333333333333, | |
| "acc_stderr": 0.013736245342311012 | |
| }, | |
| "cb": { | |
| "acc": 0.5, | |
| "acc_stderr": 0.06741998624632421, | |
| "f1": 0.3553459119496855 | |
| }, | |
| "copa": { | |
| "acc": 0.78, | |
| "acc_stderr": 0.04163331998932263 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.533559051981677, | |
| "acc_stderr": 0.0049785296421409365, | |
| "acc_norm": 0.7088229436367257, | |
| "acc_norm_stderr": 0.0045337646862119935 | |
| }, | |
| "rte": { | |
| "acc": 0.5523465703971119, | |
| "acc_stderr": 0.02993107036293953 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6479873717442778, | |
| "acc_stderr": 0.013422874824929714 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7477284874398717, | |
| "acc_stderr": 0.010043504206387305 | |
| }, | |
| "boolq": { | |
| "acc": 0.653211009174312, | |
| "acc_stderr": 0.008324380793263166 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6847643097643098, | |
| "acc_stderr": 0.009533589368505855, | |
| "acc_norm": 0.6393097643097643, | |
| "acc_norm_stderr": 0.00985351210841675 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3361774744027304, | |
| "acc_stderr": 0.013804855026205758, | |
| "acc_norm": 0.35921501706484643, | |
| "acc_norm_stderr": 0.014020224155839152 | |
| }, | |
| "sciq": { | |
| "acc": 0.93, | |
| "acc_stderr": 0.0080724943583235, | |
| "acc_norm": 0.925, | |
| "acc_norm_stderr": 0.008333333333333345 | |
| }, | |
| "piqa": { | |
| "acc": 0.7752992383025027, | |
| "acc_stderr": 0.009738282586548384, | |
| "acc_norm": 0.7856365614798694, | |
| "acc_norm_stderr": 0.00957484213605097 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |