| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.309, | |
| "acc_stderr": 0.014619600977206494 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.332, | |
| "acc_stderr": 0.014899597242811482 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.32083333333333336, | |
| "acc_stderr": 0.013480882752851557 | |
| }, | |
| "cb": { | |
| "acc": 0.5, | |
| "acc_stderr": 0.06741998624632421, | |
| "f1": 0.34848484848484845 | |
| }, | |
| "copa": { | |
| "acc": 0.66, | |
| "acc_stderr": 0.04760952285695237 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.3626767576180044, | |
| "acc_stderr": 0.004797900720081492, | |
| "acc_norm": 0.4462258514240191, | |
| "acc_norm_stderr": 0.004960839986099517 | |
| }, | |
| "rte": { | |
| "acc": 0.48375451263537905, | |
| "acc_stderr": 0.030080573208738064 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5303867403314917, | |
| "acc_stderr": 0.014026510839428751 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6317477284874399, | |
| "acc_stderr": 0.011153823258531747 | |
| }, | |
| "boolq": { | |
| "acc": 0.5486238532110091, | |
| "acc_stderr": 0.00870360524342519 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5311447811447811, | |
| "acc_stderr": 0.010239860250021748, | |
| "acc_norm": 0.5214646464646465, | |
| "acc_norm_stderr": 0.010250325159456656 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.23122866894197952, | |
| "acc_stderr": 0.012320858834772274, | |
| "acc_norm": 0.2508532423208191, | |
| "acc_norm_stderr": 0.012668198621315433 | |
| }, | |
| "sciq": { | |
| "acc": 0.886, | |
| "acc_stderr": 0.010055103435823328, | |
| "acc_norm": 0.877, | |
| "acc_norm_stderr": 0.010391293421849877 | |
| }, | |
| "piqa": { | |
| "acc": 0.6980413492927094, | |
| "acc_stderr": 0.01071173289158836, | |
| "acc_norm": 0.690968443960827, | |
| "acc_norm_stderr": 0.010781419464406979 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |