| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.332, | |
| "acc_stderr": 0.014899597242811478 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.315, | |
| "acc_stderr": 0.014696631960792492 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3233333333333333, | |
| "acc_stderr": 0.013508372867300212 | |
| }, | |
| "cb": { | |
| "acc": 0.5, | |
| "acc_stderr": 0.06741998624632421, | |
| "f1": 0.36179337231968806 | |
| }, | |
| "copa": { | |
| "acc": 0.85, | |
| "acc_stderr": 0.035887028128263734 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5313682533359888, | |
| "acc_stderr": 0.004979952166595539, | |
| "acc_norm": 0.7123083051185023, | |
| "acc_norm_stderr": 0.004517614647703246 | |
| }, | |
| "rte": { | |
| "acc": 0.5595667870036101, | |
| "acc_stderr": 0.029882123363118712 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6495659037095501, | |
| "acc_stderr": 0.013409047676670192 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7530732228754676, | |
| "acc_stderr": 0.00997199136038898 | |
| }, | |
| "boolq": { | |
| "acc": 0.6688073394495413, | |
| "acc_stderr": 0.008231583858517822 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6902356902356902, | |
| "acc_stderr": 0.00948817285190372, | |
| "acc_norm": 0.6734006734006734, | |
| "acc_norm_stderr": 0.009623047038267657 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.35921501706484643, | |
| "acc_stderr": 0.014020224155839141, | |
| "acc_norm": 0.3651877133105802, | |
| "acc_norm_stderr": 0.014070265519268804 | |
| }, | |
| "sciq": { | |
| "acc": 0.937, | |
| "acc_stderr": 0.007687007876286423, | |
| "acc_norm": 0.931, | |
| "acc_norm_stderr": 0.008018934050315158 | |
| }, | |
| "piqa": { | |
| "acc": 0.7763873775843307, | |
| "acc_stderr": 0.009721489519176297, | |
| "acc_norm": 0.7910772578890098, | |
| "acc_norm_stderr": 0.009485227030105093 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |