| { |
| "prompt": 2, |
| "orders": 2, |
| "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 held-out calibration decisions (10 family folds, each refitted without its families) have a one-sided 95% Clopper-Pearson upper bound <= max_error; a type without one defers", |
| "temperature_per_prim": { |
| "choice": 1.2968395546510096, |
| "noul": 2.3784142300054416, |
| "score": 2.1810154653305154 |
| }, |
| "temperature_per_bucket": { |
| "choice": { |
| "2-10": 1.5422108254079405, |
| "11-26": 1.189207115002721, |
| "27-255": 0.8408964152537145 |
| } |
| }, |
| "thresholds": { |
| "0.01": { |
| "confidence": null, |
| "method": "per type: the most permissive threshold whose accepted held-out calibration decisions (10 family folds, each refitted without its families) 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 held-out calibration decisions (10 family folds, each refitted without its families) have a one-sided 95% Clopper-Pearson upper bound <= max_error; a type without one defers", |
| "per_type": { |
| "choice": { |
| "confidence": 0.9324167119792968, |
| "heldout_coverage": 0.5202122283981809, |
| "heldout_error": 0.027197668771248185 |
| }, |
| "noul": { |
| "confidence": 0.8507623192282783, |
| "heldout_coverage": 0.5371807000946074, |
| "heldout_error": 0.017964071856287456 |
| }, |
| "score": { |
| "confidence": null, |
| "heldout_coverage": 0.0, |
| "heldout_error": null |
| } |
| }, |
| "heldout_coverage": 0.44773302875304166, |
| "heldout_error": 0.024290150842945835 |
| } |
| }, |
| "gates": { |
| "ece_lofo": { |
| "choice": { |
| "pooled": 0.03112408088578445, |
| "n": 11874, |
| "gate": 0.05, |
| "pass": true |
| }, |
| "noul": { |
| "pooled": 0.04447093014819661, |
| "n": 5285, |
| "gate": 0.05, |
| "pass": true |
| }, |
| "score": { |
| "pooled": 0.033942375912894035, |
| "n": 2978, |
| "gate": 0.05, |
| "pass": true |
| } |
| }, |
| "ece_worst_fold": 0.09030911583398289, |
| "worst_fold": "agent", |
| "automated_at": { |
| "0.01": 0.0, |
| "0.05": 0.44773302875304166 |
| }, |
| "automated_error": { |
| "0.01": null, |
| "0.05": 0.024290150842945835 |
| }, |
| "note": "held-out numbers: every evaluation row predicted by the fit without its family" |
| }, |
| "test": { |
| "T": { |
| "choice": { |
| "acc": 0.8319858514401213, |
| "ece": 0.030022113337473177, |
| "n": 11874 |
| }, |
| "noul": { |
| "acc": 0.8700094607379376, |
| "ece": 0.04472332079999025, |
| "n": 5285 |
| }, |
| "score": { |
| "acc": 0.6658831430490262, |
| "ece": 0.03382839088578605, |
| "n": 2978 |
| }, |
| "all": { |
| "acc": 0.8174008044892487, |
| "ece": 0.02427001174990892, |
| "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": 2, |
| "readouts": "basal-serve --mode fast, bf16" |
| }, |
| "certified_on": "the same split (held-out folds)", |
| "evaluated_on": "development and test splits, leave-one-family-out and leave-one-language-out" |
| } |
|
|
