| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.339, | |
| "acc_stderr": 0.014976758771620342 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.336, | |
| "acc_stderr": 0.014944140233795027 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3458333333333333, | |
| "acc_stderr": 0.013736245342311012 | |
| }, | |
| "cb": { | |
| "acc": 0.39285714285714285, | |
| "acc_stderr": 0.0658538889806635, | |
| "f1": 0.19047619047619047 | |
| }, | |
| "copa": { | |
| "acc": 0.82, | |
| "acc_stderr": 0.038612291966536955 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5356502688707429, | |
| "acc_stderr": 0.004977081808179436, | |
| "acc_norm": 0.7113124875522804, | |
| "acc_norm_stderr": 0.004522262128177007 | |
| }, | |
| "rte": { | |
| "acc": 0.628158844765343, | |
| "acc_stderr": 0.029091018492217447 | |
| }, | |
| "winogrande": { | |
| "acc": 0.648776637726914, | |
| "acc_stderr": 0.013415981370545126 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7589524318546232, | |
| "acc_stderr": 0.009890946490576938 | |
| }, | |
| "boolq": { | |
| "acc": 0.6351681957186545, | |
| "acc_stderr": 0.00841944098496366 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6734006734006734, | |
| "acc_stderr": 0.00962304703826764, | |
| "acc_norm": 0.5892255892255892, | |
| "acc_norm_stderr": 0.010095101349348653 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.31143344709897613, | |
| "acc_stderr": 0.013532472099850945, | |
| "acc_norm": 0.3412969283276451, | |
| "acc_norm_stderr": 0.013855831287497719 | |
| }, | |
| "sciq": { | |
| "acc": 0.896, | |
| "acc_stderr": 0.009658016218524298, | |
| "acc_norm": 0.823, | |
| "acc_norm_stderr": 0.012075463420375061 | |
| }, | |
| "piqa": { | |
| "acc": 0.780739934711643, | |
| "acc_stderr": 0.00965335746360531, | |
| "acc_norm": 0.7894450489662677, | |
| "acc_norm_stderr": 0.009512378081238743 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |