| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.328, | |
| "acc_stderr": 0.014853842487270334 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.321, | |
| "acc_stderr": 0.01477082181793464 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33416666666666667, | |
| "acc_stderr": 0.013622434813136783 | |
| }, | |
| "cb": { | |
| "acc": 0.42857142857142855, | |
| "acc_stderr": 0.06672848092813058, | |
| "f1": 0.30424242424242426 | |
| }, | |
| "copa": { | |
| "acc": 0.82, | |
| "acc_stderr": 0.038612291966536955 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.528779127663812, | |
| "acc_stderr": 0.004981509099276353, | |
| "acc_norm": 0.7054371639115714, | |
| "acc_norm_stderr": 0.004549143750428458 | |
| }, | |
| "rte": { | |
| "acc": 0.5992779783393501, | |
| "acc_stderr": 0.029497229237163147 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6464088397790055, | |
| "acc_stderr": 0.013436541262599955 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7594869053981828, | |
| "acc_stderr": 0.009883453084862687 | |
| }, | |
| "boolq": { | |
| "acc": 0.6584097859327217, | |
| "acc_stderr": 0.008294560677768487 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6914983164983165, | |
| "acc_stderr": 0.009477472342978122, | |
| "acc_norm": 0.6734006734006734, | |
| "acc_norm_stderr": 0.009623047038267656 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3455631399317406, | |
| "acc_stderr": 0.01389693846114569, | |
| "acc_norm": 0.3583617747440273, | |
| "acc_norm_stderr": 0.014012883334859859 | |
| }, | |
| "sciq": { | |
| "acc": 0.931, | |
| "acc_stderr": 0.008018934050315157, | |
| "acc_norm": 0.923, | |
| "acc_norm_stderr": 0.008434580140240634 | |
| }, | |
| "piqa": { | |
| "acc": 0.7763873775843307, | |
| "acc_stderr": 0.009721489519176289, | |
| "acc_norm": 0.7872687704026116, | |
| "acc_norm_stderr": 0.009548223123047346 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |