Implement and debug late-stage improvements for tabular ML competitions: pseudo-labeling, ensemble meta-learning, submission post-processing, and experiment tracking. Use when: adding pseudo-labeling after base models converge; building ensemble (weighted blend / LogReg stacking / dynamic gating); calibrating predictions (Platt / isotonic); clipping or constraining output values; tracking OOF scores and diagnosing OOF vs LB divergence; deciding which submission to make final. NOT for base model training, feature engineering, or hyperparameter tuning.
This skill covers the four late-stage pipeline components applied after base models converge:
Order of operations: base models → hyperparameter tuning → pseudo-labeling → ensemble → post-processing. Violating this order contaminates the OOF used for ensemble training.
va_idx[va_idx < n_train] onlyAdd calibration only as an optional post-meta pass, disabled by default. Enable only after OOF gain is confirmed.
Key danger: For ensembles of 4+ models, the ensemble already softens extremes and calibration often hurts (−0.006 OOF observed on a 4-model binary blend). Always verify on OOF before enabling.
# config.yaml