| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.327, | |
| "acc_stderr": 0.01484221315341124 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.334, | |
| "acc_stderr": 0.014922019523732953 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.345, | |
| "acc_stderr": 0.013728421539454876 | |
| }, | |
| "cb": { | |
| "acc": 0.375, | |
| "acc_stderr": 0.06527912098338669, | |
| "f1": 0.20348173515981735 | |
| }, | |
| "copa": { | |
| "acc": 0.73, | |
| "acc_stderr": 0.044619604333847394 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4531965743875722, | |
| "acc_stderr": 0.004967872475383272, | |
| "acc_norm": 0.5952997410874328, | |
| "acc_norm_stderr": 0.004898308167211839 | |
| }, | |
| "rte": { | |
| "acc": 0.555956678700361, | |
| "acc_stderr": 0.029907396333795987 | |
| }, | |
| "winogrande": { | |
| "acc": 0.580110497237569, | |
| "acc_stderr": 0.013870943986310393 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.706574024585783, | |
| "acc_stderr": 0.010529489334744471 | |
| }, | |
| "boolq": { | |
| "acc": 0.5519877675840978, | |
| "acc_stderr": 0.008697655510897231 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.577020202020202, | |
| "acc_stderr": 0.010137328382209097, | |
| "acc_norm": 0.54503367003367, | |
| "acc_norm_stderr": 0.0102180844546026 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.25426621160409557, | |
| "acc_stderr": 0.012724999945157753, | |
| "acc_norm": 0.2841296928327645, | |
| "acc_norm_stderr": 0.013179442447653887 | |
| }, | |
| "sciq": { | |
| "acc": 0.869, | |
| "acc_stderr": 0.010674874844837954, | |
| "acc_norm": 0.845, | |
| "acc_norm_stderr": 0.011450157470799464 | |
| }, | |
| "piqa": { | |
| "acc": 0.750816104461371, | |
| "acc_stderr": 0.010091882770120216, | |
| "acc_norm": 0.7616974972796517, | |
| "acc_norm_stderr": 0.009940334245876224 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |