lm1-2b8-55b-c4-repetitions
/
evaluation
/2b855b14bc4-results_lm-eval_global_step52452_2022-12-25-19-28-17.json
| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.014933117490932575 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.014933117490932572 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3466666666666667, | |
| "acc_stderr": 0.013744022550571947 | |
| }, | |
| "cb": { | |
| "acc": 0.42857142857142855, | |
| "acc_stderr": 0.06672848092813058, | |
| "f1": 0.27104247104247103 | |
| }, | |
| "copa": { | |
| "acc": 0.71, | |
| "acc_stderr": 0.045604802157206845 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4360685122485561, | |
| "acc_stderr": 0.004948824501355481, | |
| "acc_norm": 0.5651264688309102, | |
| "acc_norm_stderr": 0.004947272454226218 | |
| }, | |
| "rte": { | |
| "acc": 0.5306859205776173, | |
| "acc_stderr": 0.030039730592197812 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5595895816890292, | |
| "acc_stderr": 0.013952330311915603 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6969535008017104, | |
| "acc_stderr": 0.010627613073376715 | |
| }, | |
| "boolq": { | |
| "acc": 0.5978593272171254, | |
| "acc_stderr": 0.008575926383211254 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5711279461279462, | |
| "acc_stderr": 0.010155440652900154, | |
| "acc_norm": 0.5037878787878788, | |
| "acc_norm_stderr": 0.010259489101351842 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.257679180887372, | |
| "acc_stderr": 0.012780770562768409, | |
| "acc_norm": 0.28071672354948807, | |
| "acc_norm_stderr": 0.01313123812697558 | |
| }, | |
| "sciq": { | |
| "acc": 0.814, | |
| "acc_stderr": 0.012310790208412815, | |
| "acc_norm": 0.737, | |
| "acc_norm_stderr": 0.013929286594259734 | |
| }, | |
| "piqa": { | |
| "acc": 0.7415669205658324, | |
| "acc_stderr": 0.010213971636773326, | |
| "acc_norm": 0.7410228509249184, | |
| "acc_norm_stderr": 0.010220966031405614 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |