v1.0.0

The first stable release of Katabatic: a single interface for training, sampling, evaluating, and versioning synthetic tabular data models.

Highlights

• 15 supported generative models behind one interface. GAN-based (GANBLR, CTGAN, PATE-GAN), diffusion (TabSyn, FairTabDiffusion), language-model (GReaT, REaLTabFormer), differentially private (MST, PrivTree, PATE-GAN) and statistical baselines (ARF, KDE, Histogram, NaiveBayes, SMOTE, SynthPop). Every model trains with train(data_dir), generates rows with sample(n_samples) and reloads a trained model from its artifacts with load_from_ref(), and each has an integration test run in CI.
• Models by name. get_model("ctgan") and list_supported_models() find models through ModelRegistry. Each model installs as its own extra, e.g. pip install "katabatic[ganblr,ctgan]", and a missing dependency raises an error naming what to install. MST also needs Private-PGM, which installs separately because it isn't on PyPI.
• Six-dimension evaluation. model.evaluate(real_df) scores fidelity, utility (train on synthetic, test on real), diversity, privacy, consistency and stability, and combines them into a weighted composite. Run a subset with dimensions=, or pass your own evaluation as pipeline=, as scikit-learn takes scoring=. A dimension that fails is reported in report.errors rather than scored.
• End-to-end pipeline with versioned artifacts. TrainTestSplitPipeline splits a CSV, trains a model, generates synthetic data and evaluates it, recording the dataset split, trained model and evaluations in an artifact store so any run can be reloaded or re-evaluated. LocalArtifactStore keeps them on disk.
• Cloud artifact storage (experimental). FsspecArtifactStore keeps artifacts in S3, GCS or Azure through a local cache, so a model trained on one machine can be evaluated on another. A write that would overwrite a newer change from another machine raises ArtifactConflictError instead. Install with the artifacts-s3, artifacts-gcs or artifacts-azure extra.
• Datasets. Five benchmark datasets ship with the package (adult, car, magic, nursery and shuttle). DatasetRegistry profiles any CSV and can check it against a model's requirements, and preprocess_tabular() cleans raw CSVs before training.
• Examples and benchmarks. Notebooks for the quickstart, evaluation and cloud storage run in CI, and per-model benchmark scripts cover the five datasets.
• Experimental models (TabEBM, TabDDPM, GANBLR++, CoDi, MedGAN, TVAE-GAN, GMM and TabKDE) live in katabatic.experimental.models, outside the stability guarantee. A model is promoted to supported once it passes the promotion contract, which the test suite enforces.
• Tooling. Python 3.11 with Poetry, a Dockerfile, Makefile targets that mirror CI, and pre-commit hooks for Ruff and conventional commits.

Upgrading from 0.3.1

• Experimental models moved to katabatic.experimental.models.<name>.
• The all extra is gone: install the extras for the models you use.
• evaluate() now runs the six-dimension evaluation on every model. Model-specific scores moved to evaluate_tstr() (GANBLR, PATE-GAN), evaluate_ks() (ARF) and evaluate_loss() (TabSyn, TabDDPM).