| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.339, | |
| "acc_stderr": 0.014976758771620345 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.329, | |
| "acc_stderr": 0.014865395385928364 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33916666666666667, | |
| "acc_stderr": 0.013672343491681822 | |
| }, | |
| "cb": { | |
| "acc": 0.5178571428571429, | |
| "acc_stderr": 0.06737697508644647, | |
| "f1": 0.25267737617135205 | |
| }, | |
| "copa": { | |
| "acc": 0.7, | |
| "acc_stderr": 0.046056618647183814 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4008165704043019, | |
| "acc_stderr": 0.0048906236932436216, | |
| "acc_norm": 0.4962158932483569, | |
| "acc_norm_stderr": 0.004989638507409917 | |
| }, | |
| "rte": { | |
| "acc": 0.44765342960288806, | |
| "acc_stderr": 0.02993107036293953 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5185477505919495, | |
| "acc_stderr": 0.014042813708888378 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6483164083377873, | |
| "acc_stderr": 0.011042025772682543 | |
| }, | |
| "boolq": { | |
| "acc": 0.4782874617737003, | |
| "acc_stderr": 0.008736805647519946 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.49747474747474746, | |
| "acc_stderr": 0.010259652668783469, | |
| "acc_norm": 0.4675925925925926, | |
| "acc_norm_stderr": 0.010238210368801886 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.24658703071672355, | |
| "acc_stderr": 0.012595726268790127, | |
| "acc_norm": 0.26621160409556316, | |
| "acc_norm_stderr": 0.012915774781523223 | |
| }, | |
| "sciq": { | |
| "acc": 0.727, | |
| "acc_stderr": 0.014095022868717598, | |
| "acc_norm": 0.68, | |
| "acc_norm_stderr": 0.014758652303574874 | |
| }, | |
| "piqa": { | |
| "acc": 0.6953210010881393, | |
| "acc_stderr": 0.010738889044325161, | |
| "acc_norm": 0.7002176278563657, | |
| "acc_norm_stderr": 0.01068968696713809 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |