| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.331, | |
| "acc_stderr": 0.014888272588203938 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.324, | |
| "acc_stderr": 0.014806864733738863 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3416666666666667, | |
| "acc_stderr": 0.013696658778002519 | |
| }, | |
| "cb": { | |
| "acc": 0.14285714285714285, | |
| "acc_stderr": 0.047184161362558305, | |
| "f1": 0.1381769825918762 | |
| }, | |
| "copa": { | |
| "acc": 0.83, | |
| "acc_stderr": 0.03775251680686371 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5295757817167894, | |
| "acc_stderr": 0.004981044370530809, | |
| "acc_norm": 0.7048396733718383, | |
| "acc_norm_stderr": 0.0045518262729780596 | |
| }, | |
| "rte": { | |
| "acc": 0.49458483754512633, | |
| "acc_stderr": 0.030094698123239966 | |
| }, | |
| "winogrande": { | |
| "acc": 0.611681136543015, | |
| "acc_stderr": 0.013697456658457232 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7514698022447889, | |
| "acc_stderr": 0.009993659448666372 | |
| }, | |
| "boolq": { | |
| "acc": 0.6467889908256881, | |
| "acc_stderr": 0.008359705247064296 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.686026936026936, | |
| "acc_stderr": 0.00952324533521551, | |
| "acc_norm": 0.6628787878787878, | |
| "acc_norm_stderr": 0.009700146509130068 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3370307167235495, | |
| "acc_stderr": 0.013813476652902265, | |
| "acc_norm": 0.35665529010238906, | |
| "acc_norm_stderr": 0.013998056902620203 | |
| }, | |
| "sciq": { | |
| "acc": 0.938, | |
| "acc_stderr": 0.0076298239962803065, | |
| "acc_norm": 0.918, | |
| "acc_norm_stderr": 0.00868051561552373 | |
| }, | |
| "piqa": { | |
| "acc": 0.7742110990206746, | |
| "acc_stderr": 0.009754980670917315, | |
| "acc_norm": 0.7867247007616975, | |
| "acc_norm_stderr": 0.00955712122586134 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |