lm1-misc-pile
/
1b191b91b
/1b191b91bpile
/evaluation
/rankeval
/lm1-1b1-91b-results_lm-eval_global_step173500_2023-01-24-17-05-38_5shots.json
| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.348, | |
| "acc_stderr": 0.01507060460376841 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.348, | |
| "acc_stderr": 0.01507060460376841 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3541666666666667, | |
| "acc_stderr": 0.013811933499570963 | |
| }, | |
| "cb": { | |
| "acc": 0.625, | |
| "acc_stderr": 0.06527912098338669, | |
| "f1": 0.4297739297739298 | |
| }, | |
| "copa": { | |
| "acc": 0.67, | |
| "acc_stderr": 0.04725815626252607 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.36247759410476, | |
| "acc_stderr": 0.00479733256599006, | |
| "acc_norm": 0.4449312885879307, | |
| "acc_norm_stderr": 0.004959425421382024 | |
| }, | |
| "rte": { | |
| "acc": 0.516245487364621, | |
| "acc_stderr": 0.030080573208738064 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5351223362273086, | |
| "acc_stderr": 0.014017773120881578 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6509887760555852, | |
| "acc_stderr": 0.011022640519108536 | |
| }, | |
| "boolq": { | |
| "acc": 0.4584097859327217, | |
| "acc_stderr": 0.00871474901770989 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5732323232323232, | |
| "acc_stderr": 0.010149141043955635, | |
| "acc_norm": 0.5669191919191919, | |
| "acc_norm_stderr": 0.010167478013701789 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.25341296928327645, | |
| "acc_stderr": 0.012710896778378606, | |
| "acc_norm": 0.2815699658703072, | |
| "acc_norm_stderr": 0.013143376735009015 | |
| }, | |
| "sciq": { | |
| "acc": 0.913, | |
| "acc_stderr": 0.008916866630745904, | |
| "acc_norm": 0.916, | |
| "acc_norm_stderr": 0.00877616208949112 | |
| }, | |
| "piqa": { | |
| "acc": 0.6898803046789989, | |
| "acc_stderr": 0.01079187656684305, | |
| "acc_norm": 0.6953210010881393, | |
| "acc_norm_stderr": 0.010738889044325161 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |