| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.304, | |
| "acc_stderr": 0.014553205687950436 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.319, | |
| "acc_stderr": 0.01474640486547348 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.013630871843821465 | |
| }, | |
| "cb": { | |
| "acc": 0.35714285714285715, | |
| "acc_stderr": 0.06460957383809221, | |
| "f1": 0.20317460317460315 | |
| }, | |
| "copa": { | |
| "acc": 0.84, | |
| "acc_stderr": 0.03684529491774709 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5318661621190998, | |
| "acc_stderr": 0.004979637330230314, | |
| "acc_norm": 0.7107149970125473, | |
| "acc_norm_stderr": 0.004525037849178835 | |
| }, | |
| "rte": { | |
| "acc": 0.5776173285198556, | |
| "acc_stderr": 0.02973162264649588 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6187845303867403, | |
| "acc_stderr": 0.01365017216416031 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.757883484767504, | |
| "acc_stderr": 0.009905870033193874 | |
| }, | |
| "boolq": { | |
| "acc": 0.637308868501529, | |
| "acc_stderr": 0.008408838061823179 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6898148148148148, | |
| "acc_stderr": 0.009491721291998517, | |
| "acc_norm": 0.6734006734006734, | |
| "acc_norm_stderr": 0.00962304703826765 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3412969283276451, | |
| "acc_stderr": 0.013855831287497719, | |
| "acc_norm": 0.36006825938566556, | |
| "acc_norm_stderr": 0.014027516814585186 | |
| }, | |
| "sciq": { | |
| "acc": 0.928, | |
| "acc_stderr": 0.008178195576218681, | |
| "acc_norm": 0.921, | |
| "acc_norm_stderr": 0.008534156773333463 | |
| }, | |
| "piqa": { | |
| "acc": 0.7763873775843307, | |
| "acc_stderr": 0.009721489519176299, | |
| "acc_norm": 0.7850924918389554, | |
| "acc_norm_stderr": 0.009583665082653308 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |