| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.014933117490932572 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.352, | |
| "acc_stderr": 0.015110404505648663 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3233333333333333, | |
| "acc_stderr": 0.013508372867300215 | |
| }, | |
| "cb": { | |
| "acc": 0.4107142857142857, | |
| "acc_stderr": 0.06633634150359541, | |
| "f1": 0.3098047785547785 | |
| }, | |
| "copa": { | |
| "acc": 0.78, | |
| "acc_stderr": 0.04163331998932262 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.47849034056960765, | |
| "acc_stderr": 0.00498516207433611, | |
| "acc_norm": 0.6403106950806612, | |
| "acc_norm_stderr": 0.00478928472395585 | |
| }, | |
| "rte": { | |
| "acc": 0.4729241877256318, | |
| "acc_stderr": 0.030052303463143706 | |
| }, | |
| "winogrande": { | |
| "acc": 0.595895816890292, | |
| "acc_stderr": 0.01379161066467086 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7279529663281668, | |
| "acc_stderr": 0.01029088806087124 | |
| }, | |
| "boolq": { | |
| "acc": 0.6143730886850153, | |
| "acc_stderr": 0.008513189460768057 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6447811447811448, | |
| "acc_stderr": 0.009820245899287119, | |
| "acc_norm": 0.6195286195286195, | |
| "acc_norm_stderr": 0.009962305992058567 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.295221843003413, | |
| "acc_stderr": 0.013329750293382316, | |
| "acc_norm": 0.3046075085324232, | |
| "acc_norm_stderr": 0.013449522109932487 | |
| }, | |
| "sciq": { | |
| "acc": 0.918, | |
| "acc_stderr": 0.008680515615523705, | |
| "acc_norm": 0.902, | |
| "acc_norm_stderr": 0.009406619184621224 | |
| }, | |
| "piqa": { | |
| "acc": 0.7562568008705114, | |
| "acc_stderr": 0.010017199471500619, | |
| "acc_norm": 0.7622415669205659, | |
| "acc_norm_stderr": 0.009932525779525492 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |