| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.34, | |
| "acc_stderr": 0.014987482264363937 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.325, | |
| "acc_stderr": 0.014818724459095527 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.35583333333333333, | |
| "acc_stderr": 0.013826518748493314 | |
| }, | |
| "cb": { | |
| "acc": 0.48214285714285715, | |
| "acc_stderr": 0.0673769750864465, | |
| "f1": 0.40095238095238095 | |
| }, | |
| "copa": { | |
| "acc": 0.8, | |
| "acc_stderr": 0.040201512610368445 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4733120892252539, | |
| "acc_stderr": 0.00498266845211894, | |
| "acc_norm": 0.6218880701055567, | |
| "acc_norm_stderr": 0.0048392473326060465 | |
| }, | |
| "rte": { | |
| "acc": 0.592057761732852, | |
| "acc_stderr": 0.029581952519606193 | |
| }, | |
| "winogrande": { | |
| "acc": 0.580110497237569, | |
| "acc_stderr": 0.013870943986310393 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7012292891501871, | |
| "acc_stderr": 0.010584692134739969 | |
| }, | |
| "boolq": { | |
| "acc": 0.598776758409786, | |
| "acc_stderr": 0.008572708337178997 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.577020202020202, | |
| "acc_stderr": 0.010137328382209094, | |
| "acc_norm": 0.5315656565656566, | |
| "acc_norm_stderr": 0.010239317603199507 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2687713310580205, | |
| "acc_stderr": 0.012955065963710691, | |
| "acc_norm": 0.29948805460750855, | |
| "acc_norm_stderr": 0.013385021637313574 | |
| }, | |
| "sciq": { | |
| "acc": 0.837, | |
| "acc_stderr": 0.011686212712746839, | |
| "acc_norm": 0.788, | |
| "acc_norm_stderr": 0.012931481864938034 | |
| }, | |
| "piqa": { | |
| "acc": 0.7578890097932536, | |
| "acc_stderr": 0.009994371269104381, | |
| "acc_norm": 0.7633297062023939, | |
| "acc_norm_stderr": 0.009916841655042809 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |