| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.336, | |
| "acc_stderr": 0.01494414023379502 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.315, | |
| "acc_stderr": 0.014696631960792506 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.34, | |
| "acc_stderr": 0.0136804957257678 | |
| }, | |
| "cb": { | |
| "acc": 0.5357142857142857, | |
| "acc_stderr": 0.06724777654937658, | |
| "f1": 0.38181818181818183 | |
| }, | |
| "copa": { | |
| "acc": 0.74, | |
| "acc_stderr": 0.04408440022768077 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.48137821151165106, | |
| "acc_stderr": 0.004986319587524962, | |
| "acc_norm": 0.6344353714399522, | |
| "acc_norm_stderr": 0.004806039039008954 | |
| }, | |
| "rte": { | |
| "acc": 0.5451263537906137, | |
| "acc_stderr": 0.029973636495415252 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5974743488555643, | |
| "acc_stderr": 0.013782866831703048 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7044361304115446, | |
| "acc_stderr": 0.01055177883937378 | |
| }, | |
| "boolq": { | |
| "acc": 0.5669724770642202, | |
| "acc_stderr": 0.008666251305518059 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6220538720538721, | |
| "acc_stderr": 0.009949405744045452, | |
| "acc_norm": 0.5787037037037037, | |
| "acc_norm_stderr": 0.010131882498193127 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.29266211604095566, | |
| "acc_stderr": 0.01329591610361942, | |
| "acc_norm": 0.32849829351535836, | |
| "acc_norm_stderr": 0.013724978465537357 | |
| }, | |
| "sciq": { | |
| "acc": 0.891, | |
| "acc_stderr": 0.00985982840703719, | |
| "acc_norm": 0.871, | |
| "acc_norm_stderr": 0.010605256784796579 | |
| }, | |
| "piqa": { | |
| "acc": 0.7551686615886833, | |
| "acc_stderr": 0.010032309105568788, | |
| "acc_norm": 0.764961915125136, | |
| "acc_norm_stderr": 0.009893146688805308 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |