{ "results": { "anli_r1": { "acc": 0.344, "acc_stderr": 0.015029633724408948 }, "anli_r2": { "acc": 0.348, "acc_stderr": 0.015070604603768408 }, "anli_r3": { "acc": 0.3258333333333333, "acc_stderr": 0.013535422043417455 }, "cb": { "acc": 0.32142857142857145, "acc_stderr": 0.06297362289056341, "f1": 0.24952959907367203 }, "copa": { "acc": 0.74, "acc_stderr": 0.044084400227680794 }, "hellaswag": { "acc": 0.38040231029675364, "acc_stderr": 0.004844935327599191, "acc_norm": 0.47918741286596295, "acc_norm_stderr": 0.00498545675216101 }, "rte": { "acc": 0.4548736462093863, "acc_stderr": 0.029973636495415252 }, "winogrande": { "acc": 0.5808997632202052, "acc_stderr": 0.01386732519221011 }, "storycloze_2016": { "acc": 0.6424371993586317, "acc_stderr": 0.011083341168827792 }, "boolq": { "acc": 0.5299694189602446, "acc_stderr": 0.008729331818314897 }, "arc_easy": { "acc": 0.6069023569023569, "acc_stderr": 0.010022540618945324, "acc_norm": 0.5993265993265994, "acc_norm_stderr": 0.010055304474255579 }, "arc_challenge": { "acc": 0.2721843003412969, "acc_stderr": 0.013006600406423702, "acc_norm": 0.3037542662116041, "acc_norm_stderr": 0.013438909184778764 }, "sciq": { "acc": 0.868, "acc_stderr": 0.010709373963528042, "acc_norm": 0.849, "acc_norm_stderr": 0.011328165223341678 }, "piqa": { "acc": 0.7285092491838956, "acc_stderr": 0.010376251176596137, "acc_norm": 0.7383025027203483, "acc_norm_stderr": 0.010255630772708229 } }, "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 } }