| { |
| "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 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": 1.3542555469368927, |
| "noul": 2.4130090616010436, |
| "score": 2.3107053937445454 |
| }, |
| "temperature_per_bucket": { |
| "choice": { |
| "2-10": 1.6339154532410998, |
| "11-26": 1.2782470235604306, |
| "27-255": 0.9038571383911009 |
| } |
| }, |
| "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": 0.9343460250507047, |
| "heldout_coverage": 0.5076638032676436, |
| "heldout_error": 0.029694757796947524 |
| }, |
| "noul": { |
| "confidence": 0.8903203843605327, |
| "heldout_coverage": 0.4565752128666036, |
| "heldout_error": 0.013675922088686288 |
| }, |
| "score": { |
| "confidence": null, |
| "heldout_coverage": 0.0, |
| "heldout_error": null |
| } |
| }, |
| "heldout_coverage": 0.4191786264090977, |
| "heldout_error": 0.02511550764127468 |
| } |
| }, |
| "gates": { |
| "ece_lofo": { |
| "choice": { |
| "pooled": 0.02410796982582807, |
| "n": 11874, |
| "gate": 0.05, |
| "pass": true |
| }, |
| "noul": { |
| "pooled": 0.04203932516042977, |
| "n": 5285, |
| "gate": 0.05, |
| "pass": true |
| }, |
| "score": { |
| "pooled": 0.03959409092958942, |
| "n": 2978, |
| "gate": 0.05, |
| "pass": true |
| } |
| }, |
| "ece_worst_fold": 0.0984699552179725, |
| "worst_fold": "agent", |
| "automated_at": { |
| "0.01": 0.0, |
| "0.05": 0.4191786264090977 |
| }, |
| "automated_error": { |
| "0.01": null, |
| "0.05": 0.02511550764127468 |
| }, |
| "note": "held-out numbers: every evaluation row predicted by the fit without its family" |
| }, |
| "test": { |
| "T": { |
| "choice": { |
| "acc": 0.8283644938521139, |
| "ece": 0.022511433054648682, |
| "n": 11874 |
| }, |
| "noul": { |
| "acc": 0.868306527909177, |
| "ece": 0.04224814168670232, |
| "n": 5285 |
| }, |
| "score": { |
| "acc": 0.6638683680322364, |
| "ece": 0.039465735336162315, |
| "n": 2978 |
| }, |
| "all": { |
| "acc": 0.8145205343397726, |
| "ece": 0.02335889271209342, |
| "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, bf16" |
| }, |
| "certified_on": "the same split (held-out folds)", |
| "evaluated_on": "development and test splits, leave-one-family-out and leave-one-language-out" |
| } |
|
|
