Forge. Measure. Prove.
Kiln runs every core ML algorithm on real public data — tabular to YOLO — with one reproducible harness and honest leaderboards.
20+
Models benchmarked
6
Dataset sources
90.8%
Best CNN accuracy
4
Tracks
Four tracks, one harness
Progressive complexity — perfect for FAANG ML interviews.
Tabular
Titanic, Breast Cancer, Wine — 7 classifiers + 4 regressors via scikit-learn & Kaggle CSVs.
sklearn · Kaggle
Unsupervised
Iris clustering, Digits PCA + SVM/kNN — dimensionality reduction with OpenCV viz.
sklearn · OpenCV
Vision
Fashion-MNIST — flatten+RF baseline vs MLP vs CNN. Quantify the deep learning lift.
Keras · OpenCV
Detection
Hard Hat Workers — YOLOv8n vs YOLOv8s mAP ablation from Roboflow Universe.
Ultralytics · Roboflow
Built like internal FAANG tooling
Typed Python API, pytest + ruff CI, committed benchmark JSON, Colab GPU notebooks — not a disconnected notebook zoo.
pip install kiln-ml[dev,vision] # Run all tracks (seed=42) kiln-benchmark --track all \ --output benchmarks/results # Tabular only (~2 min) kiln-benchmark --track tabular