| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.323, | |
| "acc_stderr": 0.014794927843348635 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.346, | |
| "acc_stderr": 0.015050266127564443 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3458333333333333, | |
| "acc_stderr": 0.013736245342311012 | |
| }, | |
| "cb": { | |
| "acc": 0.4642857142857143, | |
| "acc_stderr": 0.06724777654937658, | |
| "f1": 0.28451178451178455 | |
| }, | |
| "copa": { | |
| "acc": 0.81, | |
| "acc_stderr": 0.03942772444036622 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5323640709022107, | |
| "acc_stderr": 0.004979317515432525, | |
| "acc_norm": 0.7071300537741486, | |
| "acc_norm_stderr": 0.004541492151639238 | |
| }, | |
| "rte": { | |
| "acc": 0.5848375451263538, | |
| "acc_stderr": 0.029660066290893485 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6377269139700079, | |
| "acc_stderr": 0.01350885547625251 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7498663816141101, | |
| "acc_stderr": 0.010015143382536456 | |
| }, | |
| "boolq": { | |
| "acc": 0.6470948012232416, | |
| "acc_stderr": 0.008358060743875672 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6679292929292929, | |
| "acc_stderr": 0.009663817543072703, | |
| "acc_norm": 0.5829124579124579, | |
| "acc_norm_stderr": 0.010117738967781993 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3225255972696246, | |
| "acc_stderr": 0.013659980894277378, | |
| "acc_norm": 0.3447098976109215, | |
| "acc_norm_stderr": 0.01388881628678211 | |
| }, | |
| "sciq": { | |
| "acc": 0.894, | |
| "acc_stderr": 0.009739551265785138, | |
| "acc_norm": 0.82, | |
| "acc_norm_stderr": 0.012155153135511965 | |
| }, | |
| "piqa": { | |
| "acc": 0.7747551686615887, | |
| "acc_stderr": 0.009746643471032155, | |
| "acc_norm": 0.7867247007616975, | |
| "acc_norm_stderr": 0.009557121225861342 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |