xbooster 0.2.8

v0.2.8

on Python PyPI

What's New

Fine-Tuning Support

• New finetuner module with finetune_xgb(), finetune_lgb(), finetune_cb() helpers
• FineTuneResult dataclass with tree counts and feature metadata
• Same-feature continued training and expanded-feature warm-start
• n_base_trees parameter, from_finetune_result() classmethod, summarize_score_sources() on all constructors
• TreeSource column in scorecard output (base/finetuned)

SHAP Integration

• SHAP-based scoring for XGBoost, LightGBM, and CatBoost
• Native SHAP extraction — no external shap package required
• method="shap" in predict_score() and predict_scores()

Type Checking

• Migrated from ty to mypy with custom type stubs (xgboost, catboost, lightgbm)
• Zero type: ignore comments across all source and test files

Performance

• Vectorized construct_scorecard() for XGBoost and LightGBM
• Vectorized get_leafs() and _convert_tree_to_points() for LightGBM

Other

• 144 tests passing
• Removed requirements.txt (use uv directly)

See CHANGELOG.md for full details.