| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.334, | |
| "acc_stderr": 0.014922019523732954 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.314, | |
| "acc_stderr": 0.014683991951087974 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3416666666666667, | |
| "acc_stderr": 0.013696658778002514 | |
| }, | |
| "cb": { | |
| "acc": 0.4107142857142857, | |
| "acc_stderr": 0.06633634150359541, | |
| "f1": 0.2859744990892532 | |
| }, | |
| "copa": { | |
| "acc": 0.84, | |
| "acc_stderr": 0.0368452949177471 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5349531965743876, | |
| "acc_stderr": 0.004977574188421318, | |
| "acc_norm": 0.7157936666002789, | |
| "acc_norm_stderr": 0.004501137895230712 | |
| }, | |
| "rte": { | |
| "acc": 0.5740072202166066, | |
| "acc_stderr": 0.029764956741777652 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6385161799526441, | |
| "acc_stderr": 0.013502479670791292 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.760555852485302, | |
| "acc_stderr": 0.009868402764412846 | |
| }, | |
| "boolq": { | |
| "acc": 0.6681957186544343, | |
| "acc_stderr": 0.008235412870849404 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6877104377104377, | |
| "acc_stderr": 0.009509325983631455, | |
| "acc_norm": 0.6683501683501684, | |
| "acc_norm_stderr": 0.009660733780923952 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3464163822525597, | |
| "acc_stderr": 0.013905011180063253, | |
| "acc_norm": 0.36006825938566556, | |
| "acc_norm_stderr": 0.014027516814585186 | |
| }, | |
| "sciq": { | |
| "acc": 0.942, | |
| "acc_stderr": 0.007395315455792948, | |
| "acc_norm": 0.935, | |
| "acc_norm_stderr": 0.007799733061832016 | |
| }, | |
| "piqa": { | |
| "acc": 0.7834602829162133, | |
| "acc_stderr": 0.009609984714384612, | |
| "acc_norm": 0.7867247007616975, | |
| "acc_norm_stderr": 0.00955712122586134 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |