| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.322, | |
| "acc_stderr": 0.014782913600996674 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.332, | |
| "acc_stderr": 0.014899597242811483 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3433333333333333, | |
| "acc_stderr": 0.01371263383046586 | |
| }, | |
| "cb": { | |
| "acc": 0.39285714285714285, | |
| "acc_stderr": 0.0658538889806635, | |
| "f1": 0.32857142857142857 | |
| }, | |
| "copa": { | |
| "acc": 0.82, | |
| "acc_stderr": 0.038612291966536955 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5291774546903008, | |
| "acc_stderr": 0.004981278326428018, | |
| "acc_norm": 0.7015534754033061, | |
| "acc_norm_stderr": 0.004566412808642458 | |
| }, | |
| "rte": { | |
| "acc": 0.6064981949458483, | |
| "acc_stderr": 0.029405839314203198 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6416732438831886, | |
| "acc_stderr": 0.01347658117256753 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7589524318546232, | |
| "acc_stderr": 0.00989094649057694 | |
| }, | |
| "boolq": { | |
| "acc": 0.6461773700305811, | |
| "acc_stderr": 0.008362983020904467 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6654040404040404, | |
| "acc_stderr": 0.009682137724327909, | |
| "acc_norm": 0.6595117845117845, | |
| "acc_norm_stderr": 0.009723676813825861 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3293515358361775, | |
| "acc_stderr": 0.013734057652635473, | |
| "acc_norm": 0.3515358361774744, | |
| "acc_norm_stderr": 0.013952413699600943 | |
| }, | |
| "sciq": { | |
| "acc": 0.932, | |
| "acc_stderr": 0.007964887911291605, | |
| "acc_norm": 0.918, | |
| "acc_norm_stderr": 0.008680515615523732 | |
| }, | |
| "piqa": { | |
| "acc": 0.7818280739934712, | |
| "acc_stderr": 0.009636081958374381, | |
| "acc_norm": 0.7905331882480957, | |
| "acc_norm_stderr": 0.009494302979819806 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |