|
{ |
|
"results": { |
|
"anli_r1": { |
|
"acc": 0.334, |
|
"acc_stderr": 0.014922019523732977 |
|
}, |
|
"anli_r2": { |
|
"acc": 0.351, |
|
"acc_stderr": 0.015100563798316407 |
|
}, |
|
"anli_r3": { |
|
"acc": 0.32083333333333336, |
|
"acc_stderr": 0.013480882752851543 |
|
}, |
|
"cb": { |
|
"acc": 0.42857142857142855, |
|
"acc_stderr": 0.06672848092813058, |
|
"f1": 0.2943692088382039 |
|
}, |
|
"copa": { |
|
"acc": 0.77, |
|
"acc_stderr": 0.04229525846816506 |
|
}, |
|
"hellaswag": { |
|
"acc": 0.472814180442143, |
|
"acc_stderr": 0.004982400368939667, |
|
"acc_norm": 0.6304521011750648, |
|
"acc_norm_stderr": 0.00481695881772609 |
|
}, |
|
"rte": { |
|
"acc": 0.5415162454873647, |
|
"acc_stderr": 0.029992535385373314 |
|
}, |
|
"winogrande": { |
|
"acc": 0.5911602209944752, |
|
"acc_stderr": 0.0138169542951357 |
|
}, |
|
"storycloze_2016": { |
|
"acc": 0.7188669160876536, |
|
"acc_stderr": 0.0103958360916281 |
|
}, |
|
"boolq": { |
|
"acc": 0.6073394495412844, |
|
"acc_stderr": 0.008541161248702913 |
|
}, |
|
"arc_easy": { |
|
"acc": 0.6258417508417509, |
|
"acc_stderr": 0.009929516948977627, |
|
"acc_norm": 0.6022727272727273, |
|
"acc_norm_stderr": 0.010042861602178058 |
|
}, |
|
"arc_challenge": { |
|
"acc": 0.3003412969283277, |
|
"acc_stderr": 0.013395909309956995, |
|
"acc_norm": 0.3225255972696246, |
|
"acc_norm_stderr": 0.013659980894277373 |
|
}, |
|
"sciq": { |
|
"acc": 0.914, |
|
"acc_stderr": 0.008870325962594766, |
|
"acc_norm": 0.901, |
|
"acc_norm_stderr": 0.009449248027662734 |
|
}, |
|
"piqa": { |
|
"acc": 0.7513601741022851, |
|
"acc_stderr": 0.010084511234296852, |
|
"acc_norm": 0.7687704026115343, |
|
"acc_norm_stderr": 0.009837063180625336 |
|
} |
|
}, |
|
"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 |
|
} |
|
} |