| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.338, | |
| "acc_stderr": 0.014965960710224482 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.379, | |
| "acc_stderr": 0.01534909100222535 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.35333333333333333, | |
| "acc_stderr": 0.013804572162314937 | |
| }, | |
| "cb": { | |
| "acc": 0.44642857142857145, | |
| "acc_stderr": 0.06703189227942398, | |
| "f1": 0.2940620782726046 | |
| }, | |
| "copa": { | |
| "acc": 0.82, | |
| "acc_stderr": 0.03861229196653697 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.476000796654053, | |
| "acc_stderr": 0.004984030250507291, | |
| "acc_norm": 0.6342362079267079, | |
| "acc_norm_stderr": 0.004806593424942258 | |
| }, | |
| "rte": { | |
| "acc": 0.5234657039711191, | |
| "acc_stderr": 0.03006330041190266 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6101026045777427, | |
| "acc_stderr": 0.013707547317008463 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7284874398717264, | |
| "acc_stderr": 0.010284547617192592 | |
| }, | |
| "boolq": { | |
| "acc": 0.636085626911315, | |
| "acc_stderr": 0.008414918909128852 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6460437710437711, | |
| "acc_stderr": 0.009812370644174426, | |
| "acc_norm": 0.6241582491582491, | |
| "acc_norm_stderr": 0.009938436373170616 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.29180887372013653, | |
| "acc_stderr": 0.013284525292403506, | |
| "acc_norm": 0.3054607508532423, | |
| "acc_norm_stderr": 0.013460080478002505 | |
| }, | |
| "sciq": { | |
| "acc": 0.918, | |
| "acc_stderr": 0.008680515615523725, | |
| "acc_norm": 0.914, | |
| "acc_norm_stderr": 0.008870325962594766 | |
| }, | |
| "piqa": { | |
| "acc": 0.7589771490750816, | |
| "acc_stderr": 0.009979042717267314, | |
| "acc_norm": 0.7600652883569097, | |
| "acc_norm_stderr": 0.009963625892809545 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |