| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.327, | |
| "acc_stderr": 0.014842213153411249 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.346, | |
| "acc_stderr": 0.015050266127564441 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.325, | |
| "acc_stderr": 0.013526454480351018 | |
| }, | |
| "cb": { | |
| "acc": 0.375, | |
| "acc_stderr": 0.06527912098338669, | |
| "f1": 0.2151416122004357 | |
| }, | |
| "copa": { | |
| "acc": 0.74, | |
| "acc_stderr": 0.04408440022768077 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.39494124676359293, | |
| "acc_stderr": 0.00487839022659172, | |
| "acc_norm": 0.483469428400717, | |
| "acc_norm_stderr": 0.004987053652540279 | |
| }, | |
| "rte": { | |
| "acc": 0.4657039711191336, | |
| "acc_stderr": 0.030025579819366426 | |
| }, | |
| "winogrande": { | |
| "acc": 0.49329123914759276, | |
| "acc_stderr": 0.014051220692330342 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.649919828968466, | |
| "acc_stderr": 0.011030440255782963 | |
| }, | |
| "boolq": { | |
| "acc": 0.4883792048929664, | |
| "acc_stderr": 0.008742692742551265 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.494949494949495, | |
| "acc_stderr": 0.010259260102565858, | |
| "acc_norm": 0.4515993265993266, | |
| "acc_norm_stderr": 0.010211600726405215 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.23464163822525597, | |
| "acc_stderr": 0.012383873560768675, | |
| "acc_norm": 0.2593856655290102, | |
| "acc_norm_stderr": 0.012808273573927094 | |
| }, | |
| "sciq": { | |
| "acc": 0.69, | |
| "acc_stderr": 0.014632638658632893, | |
| "acc_norm": 0.633, | |
| "acc_norm_stderr": 0.015249378464171756 | |
| }, | |
| "piqa": { | |
| "acc": 0.6931447225244831, | |
| "acc_stderr": 0.010760295070580359, | |
| "acc_norm": 0.6893362350380848, | |
| "acc_norm_stderr": 0.01079707893372768 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |