| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.326, | |
| "acc_stderr": 0.014830507204541037 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.334, | |
| "acc_stderr": 0.014922019523732961 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3258333333333333, | |
| "acc_stderr": 0.013535422043417455 | |
| }, | |
| "cb": { | |
| "acc": 0.48214285714285715, | |
| "acc_stderr": 0.0673769750864465, | |
| "f1": 0.33001107419712067 | |
| }, | |
| "copa": { | |
| "acc": 0.75, | |
| "acc_stderr": 0.04351941398892446 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4678350926110337, | |
| "acc_stderr": 0.004979446038824758, | |
| "acc_norm": 0.6130252937661821, | |
| "acc_norm_stderr": 0.0048606237334611405 | |
| }, | |
| "rte": { | |
| "acc": 0.5451263537906137, | |
| "acc_stderr": 0.029973636495415255 | |
| }, | |
| "winogrande": { | |
| "acc": 0.585635359116022, | |
| "acc_stderr": 0.013844846232268563 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7097808658471406, | |
| "acc_stderr": 0.010495529690730063 | |
| }, | |
| "boolq": { | |
| "acc": 0.6045871559633027, | |
| "acc_stderr": 0.008551600109082904 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6212121212121212, | |
| "acc_stderr": 0.009953737656542037, | |
| "acc_norm": 0.5829124579124579, | |
| "acc_norm_stderr": 0.010117738967781986 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.28242320819112626, | |
| "acc_stderr": 0.013155456884097224, | |
| "acc_norm": 0.3097269624573379, | |
| "acc_norm_stderr": 0.013512058415238361 | |
| }, | |
| "sciq": { | |
| "acc": 0.888, | |
| "acc_stderr": 0.009977753031397234, | |
| "acc_norm": 0.862, | |
| "acc_norm_stderr": 0.010912152632504401 | |
| }, | |
| "piqa": { | |
| "acc": 0.7453754080522307, | |
| "acc_stderr": 0.010164432237060492, | |
| "acc_norm": 0.7595212187159956, | |
| "acc_norm_stderr": 0.009971345364651064 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |