| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.324, | |
| "acc_stderr": 0.01480686473373886 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.329, | |
| "acc_stderr": 0.014865395385928359 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3333333333333333, | |
| "acc_stderr": 0.01361395001022561 | |
| }, | |
| "cb": { | |
| "acc": 0.3392857142857143, | |
| "acc_stderr": 0.06384226561930825, | |
| "f1": 0.33391833391833387 | |
| }, | |
| "copa": { | |
| "acc": 0.82, | |
| "acc_stderr": 0.03861229196653697 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4781915952997411, | |
| "acc_stderr": 0.004985032806802436, | |
| "acc_norm": 0.6330412268472416, | |
| "acc_norm_stderr": 0.004809901151234833 | |
| }, | |
| "rte": { | |
| "acc": 0.5379061371841155, | |
| "acc_stderr": 0.030009848912529117 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5927387529597474, | |
| "acc_stderr": 0.013808654122417845 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7295563869588455, | |
| "acc_stderr": 0.010271810373331022 | |
| }, | |
| "boolq": { | |
| "acc": 0.6311926605504588, | |
| "acc_stderr": 0.008438656079759072 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6384680134680135, | |
| "acc_stderr": 0.00985850654316206, | |
| "acc_norm": 0.625, | |
| "acc_norm_stderr": 0.009933992677987828 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.29948805460750855, | |
| "acc_stderr": 0.013385021637313565, | |
| "acc_norm": 0.3267918088737201, | |
| "acc_norm_stderr": 0.01370666597558734 | |
| }, | |
| "sciq": { | |
| "acc": 0.913, | |
| "acc_stderr": 0.008916866630745906, | |
| "acc_norm": 0.908, | |
| "acc_norm_stderr": 0.0091443763931511 | |
| }, | |
| "piqa": { | |
| "acc": 0.7568008705114254, | |
| "acc_stderr": 0.010009611953858917, | |
| "acc_norm": 0.7589771490750816, | |
| "acc_norm_stderr": 0.009979042717267315 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |