MAMUT

Machine Automated Modelling and Utility Toolkit for tabular classification.

Overview

MAMUT is a Python toolkit for transparent classification workflows on tabular data. It bundles preprocessing, Optuna-driven hyperparameter optimization, model comparison, validation diagnostics, and reporting into a single workflow built on scikit-learn, XGBoost, LightGBM, and CatBoost.

MAMUT is best used as a readable baseline and experiment report generator for beginners, small teams, and portfolio-scale projects. It is not positioned as a replacement for industrial AutoML systems such as AutoGluon, FLAML, or H2O AutoML; its value is in showing what was tried, how the result was validated, and whether simple baselines challenge the selected model.

Key Features

• End-to-end preprocessing: missing values, categorical encoding, skew correction, scaling, optional outlier filtering, imbalance handling (SMOTE/undersampling/SMOTETomek), optional feature selection, and PCA.
• Model search across common and boosted classifiers, including LogisticRegression, RandomForestClassifier, ExtraTreesClassifier, HistGradientBoostingClassifier, XGBClassifier, LGBMClassifier, CatBoostClassifier, GaussianNB, SVC, MLPClassifier, and KNeighborsClassifier.
• Hyperparameter optimization with Optuna (TPE/Bayesian or random search).
• Validation-based model selection with optional nested CV, group-disjoint splits, and final holdout evaluation.
• Evidence reporting: leakage checks, fitted-candidate comparison, dummy/logistic/random-forest baselines, and descriptive score-stability intervals from repeated stratified or grouped CV.
• Report generation via evaluate() with metrics, plots, and SHAP explanations.
• Configurable artifacts: fit() keeps models in memory by default and saves fitted models only when save_models=True.
• Reproducible benchmark diagnostics via scripts/benchmark_evidence.py.
• External Kaggle benchmark harness with immutable campaign records, explicit validation estimands, and submission controls.

Recorded Benchmark Evidence

The current evidence benchmark deliberately exposes cases where a simple baseline challenges the validation-selected candidate. On Kaggle Spaceship Titanic, the best recorded MAMUT public submission scored 0.80617 using a CatBoost-focused, post-leaderboard development campaign. This is evidence of an improved auditable baseline, not a state-of-the-art or private-leaderboard claim. Protocol details and limitations are recorded in the Kaggle benchmarks documentation.

Installation

Python 3.12 is the target runtime (see .python-version).

From PyPI:

pip install mamut

From source:

git clone https://github.com/przybytniowskaj/Mamut.git
cd Mamut
uv sync --all-groups

For development with uv:

uv sync --all-groups

Quickstart

from sklearn.datasets import load_iris
from mamut import Mamut

X, y = load_iris(as_frame=True, return_X_y=True)

mamut = Mamut(
    n_iterations=1,
    optimization_method="random_search",
    holdout_size=0.2,
    refit_final_model=True,
    random_state=42,
)
mamut.fit(X, y)

preds = mamut.predict(X.head())
proba = mamut.predict_proba(X.head())
report = mamut.evaluate(include_shap=False, write_html=False, save_plots=False)

Configuration Notes

• With preprocessing enabled (default), pass X as a pandas DataFrame and y as a Series.
• Targets must be categorical (float targets raise a ValueError).
• fit() performs a stratified train/validation split controlled by validation_size and random_state; pass groups= to keep related rows in the same partition.
• Set holdout_size or pass X_holdout/y_holdout to reserve final evaluation data that is not used for model or ensemble selection.
• Select the optimization strategy with optimization_method="bayes" or "random_search".
• Control the search budget with n_iterations.
• Choose a model pool with search_profile="quick", "balanced", or "thorough".
• Include exact models with include_models=["RandomForestClassifier", "LGBMClassifier"], or exclude models by class name with exclude_models=["SVC"].
• Use selection_strategy="nested_cv" for fold-local tuning and model comparison when runtime matters less than selection confidence.
• Control parallelism for supported estimators with n_jobs.
• preprocessing_profile="auto" lets tree and native-boosting candidates use model-aware preprocessing; set preprocessing_profile="generic_ohe" to force the legacy shared one-hot path.
• Preprocessing options are passed directly into Mamut(...) (e.g., pca=True, feature_selection=True, num_imputation="knn"). Native categorical preprocessing falls back to one-hot when PCA, feature selection, or imbalance resampling requires numeric arrays.
• Automatic outlier row removal is disabled by default; enable it explicitly with outlier_removal=True.
• Unknown categorical levels at prediction time are ignored by one-hot profiles instead of failing the prediction.
• Set verbose=True to show model-search progress logging and Optuna progress bars.
• Use save_models=True to write fitted candidate pipelines under ./fitted_models/<timestamp>/.
• Use refit_final_model=True only after you accept validation and holdout diagnostics; the final refit uses all non-holdout modeling data and never uses holdout rows.
• score_metric expects one of: accuracy, precision, recall, f1, balanced_accuracy, jaccard, roc_auc_score.
• Configure descriptive evidence stability checks with evidence_cv_splits, evidence_cv_repeats, and evidence_confidence_level; dependent CV fold intervals are not final confidence claims.

Outputs and Reports

• mamut.best_model_: validation-selected public prediction pipeline after fit; predict() returns original target labels.
• mamut.validation_summary_: per-model validation scores and timings.
• mamut.selection_summary_: model-selection diagnostics for the active selection strategy.
• mamut.fitted_preprocessors_: fitted candidate-specific preprocessors keyed by model name.
• mamut.holdout_summary_: optional final holdout diagnostic scores when holdout data is configured; use a refitted selected model for final deployment scoring.
• mamut.evidence_report_: validation integrity, evidence-guided selection, leakage checks, baseline comparison, and score stability tables generated by evaluate() or generate_evidence().
• Use generate_evidence(dataset="holdout", include_candidate_comparison=False) for a locked final confirmation report that does not score alternate MAMUT candidates on the holdout.
• mamut.optuna_studies_: Optuna studies keyed by model name.
• mamut.evaluate(): writes an HTML report to ./mamut_report/report_<timestamp>.html and stores plots in ./mamut_report/plots/. It uses holdout data automatically when available and includes evidence sections by default.
• mamut.evaluate(include_shap=False, write_html=False, save_plots=False): run lightweight evaluation without expensive SHAP or file artifacts.
• mamut.save_best_model(path): writes the best model to an existing directory as a .joblib file.

Development

uv sync --all-groups
uv run deptry .
scripts/audit_dependencies.sh
uv run pytest
uv run pre-commit run --all-files
uv run make -C docs html
uv run sphinx-build -W --keep-going -b html docs/source docs/build/html-strict
uv run python scripts/benchmark_evidence.py
uv run python scripts/benchmark_kaggle.py spaceship-titanic --campaign-id smoke --stage development --runs 1 --n-iterations 1
uv build
uv run twine check dist/*

Documentation

• Documentation site: https://mamut.readthedocs.io/en/latest/
• Quickstart: https://mamut.readthedocs.io/en/latest/quickstart.html
• User guide: https://mamut.readthedocs.io/en/latest/user_guide.html
• Reports and artifacts: https://mamut.readthedocs.io/en/latest/reports.html
• Evidence benchmark: https://mamut.readthedocs.io/en/latest/benchmark_evidence.html
• Kaggle benchmarks: https://mamut.readthedocs.io/en/latest/kaggle_benchmarks.html
• API reference: https://mamut.readthedocs.io/en/latest/mamut.html
• Notebook walkthrough: docs/source/notebooks/walkthrough.ipynb
• Kaggle notebook source: https://github.com/przybytniowskaj/Mamut/blob/main/examples/kaggle/spaceship_titanic/mamut_auditable_baseline.ipynb
• Changelog: CHANGELOG.md

License

MIT. See LICENSE.