| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.323, | |
| "acc_stderr": 0.01479492784334864 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.35, | |
| "acc_stderr": 0.015090650341444233 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3441666666666667, | |
| "acc_stderr": 0.013720551062295755 | |
| }, | |
| "cb": { | |
| "acc": 0.3392857142857143, | |
| "acc_stderr": 0.06384226561930825, | |
| "f1": 0.26343091936312274 | |
| }, | |
| "copa": { | |
| "acc": 0.72, | |
| "acc_stderr": 0.04512608598542127 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.475502887870942, | |
| "acc_stderr": 0.004983788992681198, | |
| "acc_norm": 0.6266679944234216, | |
| "acc_norm_stderr": 0.004827006520802888 | |
| }, | |
| "rte": { | |
| "acc": 0.5631768953068592, | |
| "acc_stderr": 0.02985524739031495 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5761641673243884, | |
| "acc_stderr": 0.013888492389944508 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7156600748262961, | |
| "acc_stderr": 0.01043161412866525 | |
| }, | |
| "boolq": { | |
| "acc": 0.5886850152905199, | |
| "acc_stderr": 0.008606395426309208 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5989057239057239, | |
| "acc_stderr": 0.01005705110653437, | |
| "acc_norm": 0.5425084175084175, | |
| "acc_norm_stderr": 0.010222638127749496 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2790102389078498, | |
| "acc_stderr": 0.013106784883601336, | |
| "acc_norm": 0.3046075085324232, | |
| "acc_norm_stderr": 0.01344952210993249 | |
| }, | |
| "sciq": { | |
| "acc": 0.849, | |
| "acc_stderr": 0.011328165223341671, | |
| "acc_norm": 0.757, | |
| "acc_norm_stderr": 0.013569640199177451 | |
| }, | |
| "piqa": { | |
| "acc": 0.7551686615886833, | |
| "acc_stderr": 0.010032309105568793, | |
| "acc_norm": 0.763873775843308, | |
| "acc_norm_stderr": 0.009908965890558216 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |