reports/vllm_decisions_27b.json
| 1 | { |
| 2 | "bundle": "exports/jev-judge-qwen38-27b-v0.8", |
| 3 | "split": "test_set_30k", |
| 4 | "n": 29955, |
| 5 | "load_s": 577.0090012550354, |
| 6 | "decide_s": 211.89693593978882, |
| 7 | "rows_per_s": 141.36589501469624, |
| 8 | "prompt_tok_per_s": 15076.470954293422, |
| 9 | "metrics": { |
| 10 | "n": 29955, |
| 11 | "kl": 0.01859208686817435, |
| 12 | "js": 0.004861027689099374, |
| 13 | "n_defined": 27575, |
| 14 | "top1": 0.9218494892120361, |
| 15 | "ece": 0.0010864290705572877, |
| 16 | "mce": 0.0036070922913651327, |
| 17 | "noul_brier_soft": 0.001320662068239789, |
| 18 | "noul_brier_hard": 0.0741207458019494, |
| 19 | "noul_auroc": 0.9960508485987001, |
| 20 | "score_mae_expected": 0.09772485708550437, |
| 21 | "score_rps": 0.007668479479239242, |
| 22 | "choice_top1": 0.9034896492958069, |
| 23 | "choice_kl": 0.03659439961390813 |
| 24 | }, |
| 25 | "mixed_call_s": 1.43349289894104, |
| 26 | "mixed_decisions_match": 0.003850886444511814, |
| 27 | "mixed_generation_sample": [ |
| 28 | "We need answer user's request: \"In one sentence, what is safety stock?\" Need final one sentence. Keep concise.\n</think>\n\nSafety stock is the extra inventory kep", |
| 29 | "We need answer user: \"Write a Python one-liner that reverses a string.\" Simple. Need final concise. Could provide code. Ensure one-liner. Maybe `s[::-1]`. If th" |
| 30 | ] |
| 31 | } |