| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.312, | |
| "acc_stderr": 0.014658474370509005 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.332, | |
| "acc_stderr": 0.014899597242811485 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3258333333333333, | |
| "acc_stderr": 0.013535422043417464 | |
| }, | |
| "cb": { | |
| "acc": 0.4642857142857143, | |
| "acc_stderr": 0.0672477765493766, | |
| "f1": 0.26842105263157895 | |
| }, | |
| "copa": { | |
| "acc": 0.79, | |
| "acc_stderr": 0.040936018074033256 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4759012148974308, | |
| "acc_stderr": 0.004983982396187366, | |
| "acc_norm": 0.6352320254929297, | |
| "acc_norm_stderr": 0.0048038126319949696 | |
| }, | |
| "rte": { | |
| "acc": 0.5415162454873647, | |
| "acc_stderr": 0.029992535385373314 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5974743488555643, | |
| "acc_stderr": 0.01378286683170305 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7231427044361304, | |
| "acc_stderr": 0.010347112890276924 | |
| }, | |
| "boolq": { | |
| "acc": 0.637308868501529, | |
| "acc_stderr": 0.008408838061823179 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6346801346801347, | |
| "acc_stderr": 0.009880576614806924, | |
| "acc_norm": 0.6292087542087542, | |
| "acc_norm_stderr": 0.009911292822056918 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2935153583617747, | |
| "acc_stderr": 0.013307250444941127, | |
| "acc_norm": 0.3148464163822526, | |
| "acc_norm_stderr": 0.01357265770308495 | |
| }, | |
| "sciq": { | |
| "acc": 0.914, | |
| "acc_stderr": 0.008870325962594766, | |
| "acc_norm": 0.914, | |
| "acc_norm_stderr": 0.008870325962594766 | |
| }, | |
| "piqa": { | |
| "acc": 0.763873775843308, | |
| "acc_stderr": 0.009908965890558211, | |
| "acc_norm": 0.763873775843308, | |
| "acc_norm_stderr": 0.009908965890558218 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |