| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.014933117490932575 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.334, | |
| "acc_stderr": 0.014922019523732961 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3491666666666667, | |
| "acc_stderr": 0.013767075395077249 | |
| }, | |
| "cb": { | |
| "acc": 0.39285714285714285, | |
| "acc_stderr": 0.0658538889806635, | |
| "f1": 0.23306878306878312 | |
| }, | |
| "copa": { | |
| "acc": 0.77, | |
| "acc_stderr": 0.04229525846816506 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4695279824736108, | |
| "acc_stderr": 0.0049805063294075845, | |
| "acc_norm": 0.6132244572794264, | |
| "acc_norm_stderr": 0.004860162076330956 | |
| }, | |
| "rte": { | |
| "acc": 0.5812274368231047, | |
| "acc_stderr": 0.02969666108123484 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5753749013417522, | |
| "acc_stderr": 0.013891893150264218 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.711918760021379, | |
| "acc_stderr": 0.010472537019822578 | |
| }, | |
| "boolq": { | |
| "acc": 0.5464831804281346, | |
| "acc_stderr": 0.008707182331111644 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5538720538720538, | |
| "acc_stderr": 0.01020005782876501, | |
| "acc_norm": 0.4936868686868687, | |
| "acc_norm_stderr": 0.01025896566804443 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2636518771331058, | |
| "acc_stderr": 0.012875929151297049, | |
| "acc_norm": 0.2883959044368601, | |
| "acc_norm_stderr": 0.013238394422428175 | |
| }, | |
| "sciq": { | |
| "acc": 0.82, | |
| "acc_stderr": 0.012155153135511965, | |
| "acc_norm": 0.749, | |
| "acc_norm_stderr": 0.013718133516888921 | |
| }, | |
| "piqa": { | |
| "acc": 0.73449401523395, | |
| "acc_stderr": 0.010303308653024429, | |
| "acc_norm": 0.7475516866158868, | |
| "acc_norm_stderr": 0.010135665547362354 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |