| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.368, | |
| "acc_stderr": 0.015258073561521802 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.014933117490932577 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3416666666666667, | |
| "acc_stderr": 0.013696658778002515 | |
| }, | |
| "cb": { | |
| "acc": 0.5535714285714286, | |
| "acc_stderr": 0.06703189227942395, | |
| "f1": 0.3459575611066344 | |
| }, | |
| "copa": { | |
| "acc": 0.62, | |
| "acc_stderr": 0.04878317312145633 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.2944632543318064, | |
| "acc_stderr": 0.00454869574962096, | |
| "acc_norm": 0.32423819956184025, | |
| "acc_norm_stderr": 0.0046713286732178 | |
| }, | |
| "rte": { | |
| "acc": 0.4981949458483754, | |
| "acc_stderr": 0.030096267148976633 | |
| }, | |
| "winogrande": { | |
| "acc": 0.4972375690607735, | |
| "acc_stderr": 0.014052271211616441 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.5836451095670764, | |
| "acc_stderr": 0.011399490926937006 | |
| }, | |
| "boolq": { | |
| "acc": 0.44128440366972477, | |
| "acc_stderr": 0.008684548127832634 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.4297138047138047, | |
| "acc_stderr": 0.010157908005763678, | |
| "acc_norm": 0.39941077441077444, | |
| "acc_norm_stderr": 0.010050018228742115 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.19283276450511946, | |
| "acc_stderr": 0.011529055465663338, | |
| "acc_norm": 0.22696245733788395, | |
| "acc_norm_stderr": 0.012240491536132873 | |
| }, | |
| "sciq": { | |
| "acc": 0.717, | |
| "acc_stderr": 0.014251810906481735, | |
| "acc_norm": 0.68, | |
| "acc_norm_stderr": 0.014758652303574883 | |
| }, | |
| "piqa": { | |
| "acc": 0.6278563656147987, | |
| "acc_stderr": 0.01127796831359274, | |
| "acc_norm": 0.6207834602829162, | |
| "acc_norm_stderr": 0.011320331012905077 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |