| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.325, | |
| "acc_stderr": 0.014818724459095524 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.331, | |
| "acc_stderr": 0.014888272588203928 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3225, | |
| "acc_stderr": 0.013499258621103245 | |
| }, | |
| "cb": { | |
| "acc": 0.48214285714285715, | |
| "acc_stderr": 0.0673769750864465, | |
| "f1": 0.3461538461538461 | |
| }, | |
| "copa": { | |
| "acc": 0.66, | |
| "acc_stderr": 0.04760952285695238 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.37552280422226647, | |
| "acc_stderr": 0.004832679188788792, | |
| "acc_norm": 0.4534953196574388, | |
| "acc_norm_stderr": 0.004968151878211054 | |
| }, | |
| "rte": { | |
| "acc": 0.48375451263537905, | |
| "acc_stderr": 0.030080573208738064 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5359116022099447, | |
| "acc_stderr": 0.014016193433958308 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6419027258150721, | |
| "acc_stderr": 0.011087006809925713 | |
| }, | |
| "boolq": { | |
| "acc": 0.5094801223241591, | |
| "acc_stderr": 0.008743482951361247 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.4671717171717172, | |
| "acc_stderr": 0.010237645778853848, | |
| "acc_norm": 0.4166666666666667, | |
| "acc_norm_stderr": 0.010116282977781239 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.24232081911262798, | |
| "acc_stderr": 0.012521593295800116, | |
| "acc_norm": 0.24658703071672355, | |
| "acc_norm_stderr": 0.01259572626879013 | |
| }, | |
| "sciq": { | |
| "acc": 0.703, | |
| "acc_stderr": 0.0144568322948011, | |
| "acc_norm": 0.628, | |
| "acc_norm_stderr": 0.015292149942040577 | |
| }, | |
| "piqa": { | |
| "acc": 0.6947769314472253, | |
| "acc_stderr": 0.01074426704560648, | |
| "acc_norm": 0.6947769314472253, | |
| "acc_norm_stderr": 0.01074426704560648 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |