| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.33, | |
| "acc_stderr": 0.014876872027456734 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.314, | |
| "acc_stderr": 0.014683991951087967 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3675, | |
| "acc_stderr": 0.013923529685359282 | |
| }, | |
| "cb": { | |
| "acc": 0.48214285714285715, | |
| "acc_stderr": 0.0673769750864465, | |
| "f1": 0.36000000000000004 | |
| }, | |
| "copa": { | |
| "acc": 0.76, | |
| "acc_stderr": 0.04292346959909283 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.529874526986656, | |
| "acc_stderr": 0.004980866814462756, | |
| "acc_norm": 0.7042421828321052, | |
| "acc_norm_stderr": 0.004554499409290722 | |
| }, | |
| "rte": { | |
| "acc": 0.5379061371841155, | |
| "acc_stderr": 0.030009848912529113 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6385161799526441, | |
| "acc_stderr": 0.013502479670791285 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7514698022447889, | |
| "acc_stderr": 0.009993659448666372 | |
| }, | |
| "boolq": { | |
| "acc": 0.6406727828746177, | |
| "acc_stderr": 0.00839181177040674 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6767676767676768, | |
| "acc_stderr": 0.009597218642045324, | |
| "acc_norm": 0.6439393939393939, | |
| "acc_norm_stderr": 0.009825454608416304 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3319112627986348, | |
| "acc_stderr": 0.013760988200880536, | |
| "acc_norm": 0.3515358361774744, | |
| "acc_norm_stderr": 0.013952413699600943 | |
| }, | |
| "sciq": { | |
| "acc": 0.919, | |
| "acc_stderr": 0.008632121032139985, | |
| "acc_norm": 0.915, | |
| "acc_norm_stderr": 0.00882342636694232 | |
| }, | |
| "piqa": { | |
| "acc": 0.7823721436343852, | |
| "acc_stderr": 0.009627407474840878, | |
| "acc_norm": 0.7861806311207835, | |
| "acc_norm_stderr": 0.009565994206915606 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |