| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.33, | |
| "acc_stderr": 0.014876872027456732 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.34, | |
| "acc_stderr": 0.014987482264363937 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.35, | |
| "acc_stderr": 0.013774667009018554 | |
| }, | |
| "cb": { | |
| "acc": 0.375, | |
| "acc_stderr": 0.06527912098338669, | |
| "f1": 0.33730158730158727 | |
| }, | |
| "copa": { | |
| "acc": 0.78, | |
| "acc_stderr": 0.04163331998932261 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.444035052778331, | |
| "acc_stderr": 0.004958426152481896, | |
| "acc_norm": 0.58105954989046, | |
| "acc_norm_stderr": 0.004923772581848489 | |
| }, | |
| "rte": { | |
| "acc": 0.5342960288808665, | |
| "acc_stderr": 0.030025579819366422 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5864246250986582, | |
| "acc_stderr": 0.013840971763195303 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7103153393907001, | |
| "acc_stderr": 0.01048980809194661 | |
| }, | |
| "boolq": { | |
| "acc": 0.5871559633027523, | |
| "acc_stderr": 0.00861117243047287 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6165824915824916, | |
| "acc_stderr": 0.009976995068264717, | |
| "acc_norm": 0.5917508417508418, | |
| "acc_norm_stderr": 0.010085566195791248 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.28071672354948807, | |
| "acc_stderr": 0.01313123812697558, | |
| "acc_norm": 0.30204778156996587, | |
| "acc_norm_stderr": 0.013417519144716417 | |
| }, | |
| "sciq": { | |
| "acc": 0.883, | |
| "acc_stderr": 0.010169287802713329, | |
| "acc_norm": 0.861, | |
| "acc_norm_stderr": 0.010945263761042967 | |
| }, | |
| "piqa": { | |
| "acc": 0.73449401523395, | |
| "acc_stderr": 0.010303308653024427, | |
| "acc_norm": 0.7372143634385201, | |
| "acc_norm_stderr": 0.010269354068140777 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |