| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.014933117490932577 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.337, | |
| "acc_stderr": 0.01495508791865361 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.35083333333333333, | |
| "acc_stderr": 0.013782212417178199 | |
| }, | |
| "cb": { | |
| "acc": 0.4107142857142857, | |
| "acc_stderr": 0.06633634150359541, | |
| "f1": 0.28810120539443845 | |
| }, | |
| "copa": { | |
| "acc": 0.8, | |
| "acc_stderr": 0.040201512610368445 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4658434574785899, | |
| "acc_stderr": 0.004978124945759844, | |
| "acc_norm": 0.6102370045807608, | |
| "acc_norm_stderr": 0.0048669971103881965 | |
| }, | |
| "rte": { | |
| "acc": 0.5776173285198556, | |
| "acc_stderr": 0.029731622646495887 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5935280189423836, | |
| "acc_stderr": 0.013804448697753376 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.729021913415286, | |
| "acc_stderr": 0.010278188399635051 | |
| }, | |
| "boolq": { | |
| "acc": 0.6134556574923548, | |
| "acc_stderr": 0.008516943934341978 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5951178451178452, | |
| "acc_stderr": 0.010072423960395701, | |
| "acc_norm": 0.5585016835016835, | |
| "acc_norm_stderr": 0.010189314382749934 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.27474402730375425, | |
| "acc_stderr": 0.013044617212771227, | |
| "acc_norm": 0.3165529010238908, | |
| "acc_norm_stderr": 0.01359243151906808 | |
| }, | |
| "sciq": { | |
| "acc": 0.855, | |
| "acc_stderr": 0.011139977517890134, | |
| "acc_norm": 0.794, | |
| "acc_norm_stderr": 0.012795613612786548 | |
| }, | |
| "piqa": { | |
| "acc": 0.7459194776931447, | |
| "acc_stderr": 0.010157271999135041, | |
| "acc_norm": 0.7595212187159956, | |
| "acc_norm_stderr": 0.009971345364651068 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |