ModelForge

ModelForge is a transparent, local-first AutoML framework for discovering, evaluating, explaining, and persisting scikit-learn pipelines. It keeps the workflow inspectable: data audits, preprocessing, model screening, cross-validation, ranking, reproducibility metadata, and artifacts are all available as Python objects and local files.

Quick start

python -m pip install -e ".[dev]"
modelforge train --data examples/house_prices.csv --target price \
	--task-type regression --output model.joblib
modelforge predict --model model.joblib --data examples/house_prices_new.csv

The same workflow is available from Python:

from modelforge import AutoML

automl = AutoML(cv=5, random_state=42)
result = automl.fit(data, target="price", task_type="regression")
automl.save("artifacts/house-prices.joblib", overwrite=True)
print(result["best_model"])

What is included

ModelForge currently provides ingestion, target detection, profiling, column intelligence, data-quality and leakage auditing, preprocessing, feature engineering and selection, a model registry, screening, cross-validation, ranking, explainability, persistence, prediction validation, optimization, experiment tracking, run management, configuration, CLI workflows, failure isolation, robustness checks, and reproducibility snapshots.

Optional model and optimization integrations are declared in pyproject.toml:

python -m pip install -e ".[boosting,optimization]"

CLI

modelforge --help
modelforge models --task-type classification
modelforge experiments list --directory .modelforge/experiments
modelforge experiments get EXPERIMENT_ID
modelforge experiments compare ID_ONE ID_TWO

The CLI uses Rich tables and keeps generated models, experiment JSON, and metadata local by default.

Examples and benchmarks

• examples/end_to_end_regression.py demonstrates a complete Python workflow.
• examples/end_to_end_classification.py demonstrates classification and probability prediction.
• benchmarks/benchmark_baselines.py compares ModelForge with simple baseline estimators and records runtime and quality metrics.

Run an example with python examples/end_to_end_regression.py or run the benchmark with python benchmarks/benchmark_baselines.py.

Development

.venv\\Scripts\\python.exe -m pytest -q
ruff check .

GitHub Actions runs the test suite on supported Python versions. The roadmap for the remaining product work lives in docs/ROADMAP.md; it separates implemented foundations from integrations that require optional dependencies or additional design work.