| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.327, | |
| "acc_stderr": 0.014842213153411244 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.327, | |
| "acc_stderr": 0.014842213153411245 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.32166666666666666, | |
| "acc_stderr": 0.013490095282989521 | |
| }, | |
| "cb": { | |
| "acc": 0.5178571428571429, | |
| "acc_stderr": 0.06737697508644647, | |
| "f1": 0.32323232323232315 | |
| }, | |
| "copa": { | |
| "acc": 0.77, | |
| "acc_stderr": 0.04229525846816505 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.45210117506472813, | |
| "acc_stderr": 0.004966832553245033, | |
| "acc_norm": 0.5958972316271659, | |
| "acc_norm_stderr": 0.004897146690596266 | |
| }, | |
| "rte": { | |
| "acc": 0.48736462093862815, | |
| "acc_stderr": 0.030086851767188564 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5714285714285714, | |
| "acc_stderr": 0.013908353814606696 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6974879743452699, | |
| "acc_stderr": 0.010622307774396942 | |
| }, | |
| "boolq": { | |
| "acc": 0.618960244648318, | |
| "acc_stderr": 0.00849393752443934 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6452020202020202, | |
| "acc_stderr": 0.009817629113069696, | |
| "acc_norm": 0.6388888888888888, | |
| "acc_norm_stderr": 0.00985601342581124 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.295221843003413, | |
| "acc_stderr": 0.013329750293382318, | |
| "acc_norm": 0.32849829351535836, | |
| "acc_norm_stderr": 0.013724978465537378 | |
| }, | |
| "sciq": { | |
| "acc": 0.922, | |
| "acc_stderr": 0.008484573530118585, | |
| "acc_norm": 0.924, | |
| "acc_norm_stderr": 0.008384169266796387 | |
| }, | |
| "piqa": { | |
| "acc": 0.7399347116430903, | |
| "acc_stderr": 0.010234893249061301, | |
| "acc_norm": 0.7453754080522307, | |
| "acc_norm_stderr": 0.0101644322370605 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |