| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.31, | |
| "acc_stderr": 0.014632638658632902 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.31, | |
| "acc_stderr": 0.014632638658632905 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3283333333333333, | |
| "acc_stderr": 0.013562032919529017 | |
| }, | |
| "cb": { | |
| "acc": 0.3392857142857143, | |
| "acc_stderr": 0.06384226561930825, | |
| "f1": 0.29749748849204566 | |
| }, | |
| "copa": { | |
| "acc": 0.79, | |
| "acc_stderr": 0.040936018074033256 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4803823939454292, | |
| "acc_stderr": 0.004985939292819582, | |
| "acc_norm": 0.6294562836088429, | |
| "acc_norm_stderr": 0.004819633668832538 | |
| }, | |
| "rte": { | |
| "acc": 0.44765342960288806, | |
| "acc_stderr": 0.02993107036293953 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5887924230465666, | |
| "acc_stderr": 0.013829128358676874 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7049706039551042, | |
| "acc_stderr": 0.010546232606962289 | |
| }, | |
| "boolq": { | |
| "acc": 0.5522935779816514, | |
| "acc_stderr": 0.008697094687974059 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6262626262626263, | |
| "acc_stderr": 0.009927267058259621, | |
| "acc_norm": 0.5934343434343434, | |
| "acc_norm_stderr": 0.010079056419223527 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2883959044368601, | |
| "acc_stderr": 0.013238394422428173, | |
| "acc_norm": 0.3148464163822526, | |
| "acc_norm_stderr": 0.01357265770308495 | |
| }, | |
| "sciq": { | |
| "acc": 0.892, | |
| "acc_stderr": 0.0098200016513457, | |
| "acc_norm": 0.869, | |
| "acc_norm_stderr": 0.010674874844837954 | |
| }, | |
| "piqa": { | |
| "acc": 0.7486398258977149, | |
| "acc_stderr": 0.010121156016819259, | |
| "acc_norm": 0.7633297062023939, | |
| "acc_norm_stderr": 0.009916841655042809 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |