| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.327, | |
| "acc_stderr": 0.014842213153411249 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.35, | |
| "acc_stderr": 0.015090650341444233 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.325, | |
| "acc_stderr": 0.013526454480351025 | |
| }, | |
| "cb": { | |
| "acc": 0.4107142857142857, | |
| "acc_stderr": 0.06633634150359541, | |
| "f1": 0.26894586894586897 | |
| }, | |
| "copa": { | |
| "acc": 0.82, | |
| "acc_stderr": 0.038612291966536955 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4736108344951205, | |
| "acc_stderr": 0.0049828269166871525, | |
| "acc_norm": 0.6298546106353317, | |
| "acc_norm_stderr": 0.004818566366066934 | |
| }, | |
| "rte": { | |
| "acc": 0.5487364620938628, | |
| "acc_stderr": 0.029953149241808943 | |
| }, | |
| "winogrande": { | |
| "acc": 0.585635359116022, | |
| "acc_stderr": 0.013844846232268563 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7188669160876536, | |
| "acc_stderr": 0.010395836091628108 | |
| }, | |
| "boolq": { | |
| "acc": 0.6342507645259939, | |
| "acc_stderr": 0.00842393006885078 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6372053872053872, | |
| "acc_stderr": 0.009865936757013938, | |
| "acc_norm": 0.6077441077441077, | |
| "acc_norm_stderr": 0.010018744689650043 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2841296928327645, | |
| "acc_stderr": 0.013179442447653886, | |
| "acc_norm": 0.30802047781569963, | |
| "acc_norm_stderr": 0.013491429517292038 | |
| }, | |
| "sciq": { | |
| "acc": 0.911, | |
| "acc_stderr": 0.009008893392651528, | |
| "acc_norm": 0.885, | |
| "acc_norm_stderr": 0.010093407594904638 | |
| }, | |
| "piqa": { | |
| "acc": 0.7611534276387377, | |
| "acc_stderr": 0.0099481203853375, | |
| "acc_norm": 0.7584330794341676, | |
| "acc_norm_stderr": 0.009986718001804451 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |