| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.333, | |
| "acc_stderr": 0.014910846164229863 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.35, | |
| "acc_stderr": 0.015090650341444231 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3491666666666667, | |
| "acc_stderr": 0.013767075395077249 | |
| }, | |
| "cb": { | |
| "acc": 0.48214285714285715, | |
| "acc_stderr": 0.0673769750864465, | |
| "f1": 0.3432539682539683 | |
| }, | |
| "copa": { | |
| "acc": 0.75, | |
| "acc_stderr": 0.04351941398892446 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.45757817167894843, | |
| "acc_stderr": 0.004971789638563322, | |
| "acc_norm": 0.5881298546106354, | |
| "acc_norm_stderr": 0.004911659884506154 | |
| }, | |
| "rte": { | |
| "acc": 0.48736462093862815, | |
| "acc_stderr": 0.030086851767188564 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5580110497237569, | |
| "acc_stderr": 0.01395758407910899 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6900053447354356, | |
| "acc_stderr": 0.010695042806212553 | |
| }, | |
| "boolq": { | |
| "acc": 0.5504587155963303, | |
| "acc_stderr": 0.008700409761350798 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5361952861952862, | |
| "acc_stderr": 0.010232865550346727, | |
| "acc_norm": 0.4823232323232323, | |
| "acc_norm_stderr": 0.010253369805698971 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2636518771331058, | |
| "acc_stderr": 0.012875929151297056, | |
| "acc_norm": 0.30631399317406144, | |
| "acc_norm_stderr": 0.013470584417276513 | |
| }, | |
| "sciq": { | |
| "acc": 0.765, | |
| "acc_stderr": 0.013414729030247116, | |
| "acc_norm": 0.68, | |
| "acc_norm_stderr": 0.014758652303574885 | |
| }, | |
| "piqa": { | |
| "acc": 0.7334058759521219, | |
| "acc_stderr": 0.010316749863541369, | |
| "acc_norm": 0.733949945593036, | |
| "acc_norm_stderr": 0.010310039263352827 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |