| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.303, | |
| "acc_stderr": 0.014539683710535264 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.312, | |
| "acc_stderr": 0.01465847437050901 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3491666666666667, | |
| "acc_stderr": 0.013767075395077247 | |
| }, | |
| "cb": { | |
| "acc": 0.30357142857142855, | |
| "acc_stderr": 0.06199938655510753, | |
| "f1": 0.2927120669056153 | |
| }, | |
| "copa": { | |
| "acc": 0.79, | |
| "acc_stderr": 0.040936018074033256 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4506074487153953, | |
| "acc_stderr": 0.0049653753416431376, | |
| "acc_norm": 0.5834495120493925, | |
| "acc_norm_stderr": 0.004919794704673269 | |
| }, | |
| "rte": { | |
| "acc": 0.5342960288808665, | |
| "acc_stderr": 0.030025579819366422 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6077348066298343, | |
| "acc_stderr": 0.013722400462000883 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6996258685195083, | |
| "acc_stderr": 0.010600915927985033 | |
| }, | |
| "boolq": { | |
| "acc": 0.6134556574923548, | |
| "acc_stderr": 0.008516943934341973 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5231481481481481, | |
| "acc_stderr": 0.010248782484554473, | |
| "acc_norm": 0.4819023569023569, | |
| "acc_norm_stderr": 0.010253060653479177 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.23122866894197952, | |
| "acc_stderr": 0.012320858834772273, | |
| "acc_norm": 0.2619453924914676, | |
| "acc_norm_stderr": 0.012849054826858115 | |
| }, | |
| "sciq": { | |
| "acc": 0.88, | |
| "acc_stderr": 0.010281328012747391, | |
| "acc_norm": 0.863, | |
| "acc_norm_stderr": 0.010878848714333327 | |
| }, | |
| "piqa": { | |
| "acc": 0.6887921653971708, | |
| "acc_stderr": 0.010802263878045844, | |
| "acc_norm": 0.6866158868335147, | |
| "acc_norm_stderr": 0.010822829929195489 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |