What problem does it solve? Machine learning results on the PIDD diabetes dataset are frequently inflated by data leakage: SMOTE applied before train/test splitting, global imputation, and preprocessing outside cross-validation folds produce misleading metrics like F1 > 0.90 when the true ceiling is around 0.71. This Skill audits such pipelines, identifies leakage sources, and enforces correct methodology. ## Core Features & Use Cases - Leakage Detection: Flags severe leakage (SMOTE before splitting, causing Recall collapse to ~0.50 and Precision inflation to 1.00) and minor leakage (global imputation, ~+0.6% F1 inflation). - Correct Pipeline Enforcement: Requires imblearn.Pipeline instead of sklearn.Pipeline so SMOTE and imputation run inside each CV fold. - Environment & Citation Rules: Mandates an isolated venv with numpy>=2.0 for SHAP analysis and requires co-citing Shearer (2000) or Wirth & Hipp (2000) whenever CRISP-DM is referenced. - Use Case: Given a 10-fold CV script on PIDD (768 samples) reporting F1=0.93, the audit identifies fold-external SMOTE, rebuilds the pipeline with imblearn, and reruns to confirm F1 falls back to the ~0.71 range, producing cv_results.csv and comprehensive_results.json. ## Quick Start Audit my PIDD cross-validation script for data leakage and rerun it with preprocessing isolated inside the folds.
