| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.342, | |
| "acc_stderr": 0.01500870618212173 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.343, | |
| "acc_stderr": 0.015019206922356953 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3491666666666667, | |
| "acc_stderr": 0.013767075395077249 | |
| }, | |
| "cb": { | |
| "acc": 0.3392857142857143, | |
| "acc_stderr": 0.06384226561930825, | |
| "f1": 0.24338624338624337 | |
| }, | |
| "copa": { | |
| "acc": 0.78, | |
| "acc_stderr": 0.04163331998932261 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4737104162517427, | |
| "acc_stderr": 0.004982879340691403, | |
| "acc_norm": 0.616211909978092, | |
| "acc_norm_stderr": 0.004853134271547751 | |
| }, | |
| "rte": { | |
| "acc": 0.555956678700361, | |
| "acc_stderr": 0.029907396333795994 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5943172849250198, | |
| "acc_stderr": 0.01380020633601421 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7076429716729022, | |
| "acc_stderr": 0.010518239729787743 | |
| }, | |
| "boolq": { | |
| "acc": 0.5770642201834862, | |
| "acc_stderr": 0.008640558744656426 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5526094276094277, | |
| "acc_stderr": 0.010202832385415644, | |
| "acc_norm": 0.5004208754208754, | |
| "acc_norm_stderr": 0.010259779886094424 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2713310580204778, | |
| "acc_stderr": 0.012993807727545797, | |
| "acc_norm": 0.2960750853242321, | |
| "acc_norm_stderr": 0.013340916085246268 | |
| }, | |
| "sciq": { | |
| "acc": 0.84, | |
| "acc_stderr": 0.011598902298689005, | |
| "acc_norm": 0.758, | |
| "acc_norm_stderr": 0.013550631705555954 | |
| }, | |
| "piqa": { | |
| "acc": 0.7393906420021763, | |
| "acc_stderr": 0.010241826155811625, | |
| "acc_norm": 0.749183895538629, | |
| "acc_norm_stderr": 0.010113869547069046 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |