| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.361, | |
| "acc_stderr": 0.015195720118175124 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.332, | |
| "acc_stderr": 0.014899597242811475 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3441666666666667, | |
| "acc_stderr": 0.013720551062295755 | |
| }, | |
| "cb": { | |
| "acc": 0.44642857142857145, | |
| "acc_stderr": 0.06703189227942398, | |
| "f1": 0.3336203597397627 | |
| }, | |
| "copa": { | |
| "acc": 0.77, | |
| "acc_stderr": 0.04229525846816506 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4676359290977893, | |
| "acc_stderr": 0.004979317515432522, | |
| "acc_norm": 0.6258713403704441, | |
| "acc_norm_stderr": 0.004829081532826523 | |
| }, | |
| "rte": { | |
| "acc": 0.5054151624548736, | |
| "acc_stderr": 0.030094698123239966 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5816890292028414, | |
| "acc_stderr": 0.013863669961195908 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.721004810261892, | |
| "acc_stderr": 0.010371620932652793 | |
| }, | |
| "boolq": { | |
| "acc": 0.6152905198776758, | |
| "acc_stderr": 0.008509403073229692 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6245791245791246, | |
| "acc_stderr": 0.0099362185271143, | |
| "acc_norm": 0.6199494949494949, | |
| "acc_norm_stderr": 0.009960175831493131 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.30631399317406144, | |
| "acc_stderr": 0.013470584417276513, | |
| "acc_norm": 0.32337883959044367, | |
| "acc_norm_stderr": 0.013669421630012122 | |
| }, | |
| "sciq": { | |
| "acc": 0.913, | |
| "acc_stderr": 0.008916866630745913, | |
| "acc_norm": 0.904, | |
| "acc_norm_stderr": 0.009320454434783217 | |
| }, | |
| "piqa": { | |
| "acc": 0.7529923830250272, | |
| "acc_stderr": 0.010062268140772629, | |
| "acc_norm": 0.7540805223068553, | |
| "acc_norm_stderr": 0.010047331865625184 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |