| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.334, | |
| "acc_stderr": 0.01492201952373296 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.341, | |
| "acc_stderr": 0.0149981313484027 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3358333333333333, | |
| "acc_stderr": 0.013639261190932882 | |
| }, | |
| "cb": { | |
| "acc": 0.39285714285714285, | |
| "acc_stderr": 0.0658538889806635, | |
| "f1": 0.18803418803418803 | |
| }, | |
| "copa": { | |
| "acc": 0.82, | |
| "acc_stderr": 0.038612291966536955 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5300736904999004, | |
| "acc_stderr": 0.004980747448813311, | |
| "acc_norm": 0.7024497112129058, | |
| "acc_norm_stderr": 0.004562462665505219 | |
| }, | |
| "rte": { | |
| "acc": 0.5667870036101083, | |
| "acc_stderr": 0.029826764082138274 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6227308602999211, | |
| "acc_stderr": 0.013622567928799503 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7541421699625869, | |
| "acc_stderr": 0.009957443066942233 | |
| }, | |
| "boolq": { | |
| "acc": 0.5951070336391437, | |
| "acc_stderr": 0.008585393347962317 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6599326599326599, | |
| "acc_stderr": 0.009720765494805276, | |
| "acc_norm": 0.5984848484848485, | |
| "acc_norm_stderr": 0.010058790020755562 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3148464163822526, | |
| "acc_stderr": 0.013572657703084948, | |
| "acc_norm": 0.34897610921501704, | |
| "acc_norm_stderr": 0.013928933461382496 | |
| }, | |
| "sciq": { | |
| "acc": 0.894, | |
| "acc_stderr": 0.009739551265785141, | |
| "acc_norm": 0.828, | |
| "acc_norm_stderr": 0.011939788882495321 | |
| }, | |
| "piqa": { | |
| "acc": 0.779107725788901, | |
| "acc_stderr": 0.009679088048842217, | |
| "acc_norm": 0.7878128400435256, | |
| "acc_norm_stderr": 0.009539299828174044 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |