| { |
| "prompt": 2, |
| "orders": 1, |
| "method": "temperature per question type and per option-count bucket (choice), by NLL, fitted on the out-of-distribution calibration split only; thresholds per type: the most permissive threshold whose accepted calibration decisions have a one-sided 95% Clopper-Pearson upper bound <= max_error; a type without one defers", |
| "temperature_per_prim": { |
| "choice": 1.3739536474580891, |
| "noul": 2.9113063656843745, |
| "score": 2.4130090616010436 |
| }, |
| "temperature_per_bucket": { |
| "choice": { |
| "2-10": 1.7817974362806785, |
| "11-26": 1.189207115002721, |
| "27-255": 1.0442737824274138 |
| } |
| }, |
| "thresholds": { |
| "0.01": { |
| "confidence": null, |
| "method": "per type: the most permissive threshold whose accepted calibration decisions have a one-sided 95% Clopper-Pearson upper bound <= max_error; a type without one defers", |
| "per_type": { |
| "choice": { |
| "confidence": null, |
| "heldout_coverage": 0.0, |
| "heldout_error": null |
| }, |
| "noul": { |
| "confidence": null, |
| "heldout_coverage": 0.0, |
| "heldout_error": null |
| }, |
| "score": { |
| "confidence": null, |
| "heldout_coverage": 0.0, |
| "heldout_error": null |
| } |
| }, |
| "heldout_coverage": 0.0, |
| "heldout_error": null |
| }, |
| "0.05": { |
| "confidence": null, |
| "method": "per type: the most permissive threshold whose accepted calibration decisions have a one-sided 95% Clopper-Pearson upper bound <= max_error; a type without one defers", |
| "per_type": { |
| "choice": { |
| "confidence": 0.9286868153250845, |
| "heldout_coverage": 0.3958228061310426, |
| "heldout_error": 0.024042553191489402 |
| }, |
| "noul": { |
| "confidence": 0.8781293507775572, |
| "heldout_coverage": 0.3791863765373699, |
| "heldout_error": 0.025449101796407136 |
| }, |
| "score": { |
| "confidence": null, |
| "heldout_coverage": 0.010745466756212223, |
| "heldout_error": 0.21875 |
| } |
| }, |
| "heldout_coverage": 0.33450861598053333, |
| "heldout_error": 0.025385985748218487 |
| } |
| }, |
| "gates": { |
| "ece_lofo": { |
| "choice": { |
| "pooled": 0.014017515750226483, |
| "n": 11874, |
| "gate": 0.05, |
| "pass": true |
| }, |
| "noul": { |
| "pooled": 0.036824358355951686, |
| "n": 5285, |
| "gate": 0.05, |
| "pass": true |
| }, |
| "score": { |
| "pooled": 0.03437115639927046, |
| "n": 2978, |
| "gate": 0.05, |
| "pass": true |
| } |
| }, |
| "ece_worst_fold": 0.12806117595026437, |
| "worst_fold": "agent", |
| "automated_at": { |
| "0.01": 0.0, |
| "0.05": 0.33450861598053333 |
| }, |
| "automated_error": { |
| "0.01": null, |
| "0.05": 0.025385985748218487 |
| }, |
| "note": "held-out numbers: every evaluation row predicted by the fit without its family" |
| }, |
| "test": { |
| "T": { |
| "choice": { |
| "acc": 0.7825501094829038, |
| "ece": 0.014634286980745098, |
| "n": 11874 |
| }, |
| "noul": { |
| "acc": 0.8297067171239356, |
| "ece": 0.03698786735263829, |
| "n": 5285 |
| }, |
| "score": { |
| "acc": 0.6064472800537274, |
| "ece": 0.036633175755623124, |
| "n": 2978 |
| }, |
| "all": { |
| "acc": 0.7688831504196255, |
| "ece": 0.021431324576755978, |
| "n": 20137 |
| } |
| }, |
| "note": "evaluation pool (development and test splits), full fit" |
| }, |
| "n": { |
| "fit": 3560, |
| "many": 1500, |
| "eval": 20137 |
| }, |
| "engine": "fast", |
| "fitted_on": { |
| "split": "out-of-distribution calibration split (sealed-seed BH-, JDI- and agent-shaped families, many-option items)", |
| "orders": 1, |
| "readouts": "basal-serve --mode fast, ModelOpt FP8 package (recipe b2)" |
| }, |
| "certified_on": "the same split", |
| "evaluated_on": "development and test splits, leave-one-family-out and leave-one-language-out" |
| } |
|
|
