Croissant Baker

Automatically generate Croissant JSON-LD metadata for ML datasets — e.g. for PhysioNet, NeurIPS Datasets & Benchmarks submissions, or any platform that benefits from standardized dataset metadata.

📖 Documentation

---

Installation

pip install croissant-baker

or with uv:

uv add croissant-baker

Quick start

croissant-baker \
  --input /path/to/dataset \
  --creator "Your Name,you@example.com" \
  --description "My ML dataset" \
  --license "CC-BY-4.0" \
  --output my-dataset-croissant.jsonld

Or try with the bundled MIMIC-IV Demo test data:

git clone https://github.com/MIT-LCP/croissant-baker.git && cd croissant-baker
uv sync --group dev
croissant-baker \
  --input tests/data/input/mimiciv_demo/physionet.org/files/mimic-iv-demo/ \
  --creator "Alistair Johnson,aewj@mit.edu,https://physionet.org/" \
  --creator "Tom Pollard,tpollard@mit.edu,https://physionet.org/" \
  --name "MIMIC-IV Clinical Database Demo" \
  --description "Demo subset of MIMIC-IV containing 100 de-identified patients from Beth Israel Deaconess Medical Center" \
  --url "https://physionet.org/content/mimic-iv-demo/2.2/" \
  --license "https://opendatacommons.org/licenses/odbl/1-0/" \
  --rai-data-biases "Single-site cohort from a US academic medical centre" \
  --rai-data-limitations "Demo subset limited to 100 patients" \
  --output mimic-iv-demo-croissant.jsonld
croissant-baker validate mimic-iv-demo-croissant.jsonld

Supported formats

Format	Extensions	Notes
CSV / TSV	.csv, .tsv	Streaming with automatic type inference
Parquet	.parquet	Partitioned datasets supported
FHIR	.ndjson, .json (Bundle)	NDJSON bulk export and JSON Bundle
JSON / JSONL	.json, .jsonl	Arrays, single objects, and JSON Lines
WFDB	.hea + .dat / .atr	PhysioNet waveform data
Images	.png, .jpg, .tiff, .btf, .bmp, .gif, .webp	Dimensions and format; BigTIFF via tifffile; OME-XML header fields for OME-TIFF
DICOM	.dcm, .dicom	Modality, geometry, study/series UIDs via pydicom (header only)
NIfTI	.nii	Spatial dims, voxel spacing, TR for fMRI via nibabel (header only)
GEO SOFT	.soft	NCBI GEO family exports: attribute names, sample characteristic keys, data table columns
HDF5	.h5, .h5ad, .hdf5	Dataset paths, dtypes and shapes via h5py (structure only); AnnData and 10x table columns where the layout is recognised

Any of these may arrive wrapped in .gz, .bz2 or .xz — compression is resolved before the format is read, so cells.parquet.gz is described exactly as cells.parquet is. WFDB is the exception: a record spans several files located by path, so a compressed .hea is reported rather than described. .zip, .tar and .tgz are reported as archives, not opened. Files nothing describes are reported with a reason rather than skipped in silence.

Key features

• Automatic type inference for all supported formats
• RAI metadata via --rai-* CLI flags or --rai-config rai.yaml
• Validation against the Croissant spec via mlcroissant
• Dry-run mode, include/exclude glob filters, multiple creators

See the documentation for full CLI reference, examples, and RAI configuration.

Contributing

See CONTRIBUTING.md for guidelines and DEVELOPMENT.md for setup, testing, releases, and how to add new file handlers.

Citation

If you use Croissant Baker in your research, please cite our arXiv preprint:

@misc{attrach2026croissantbakermetadatageneration,
      title={Croissant Baker: Metadata Generation for Discoverable, Governable, and Reusable ML Datasets},
      author={Rafi Al Attrach and Rajna Fani and Sebastian Lobentanzer and Joan Giner-Miguelez and Debanshu Das and Varuni H. K. and Nobin Sarwar and Rajat Ghosh and Anwai Archit and Surbhi Motghare and Christina Conrad Parry and Luis Oala and Lara Grosso and Joaquin Vanschoren and Steffen Vogler and Sujata Goswami and Eric S. Rosenthal and Marzyeh Ghassemi and Matthew McDermott and Tom Pollard},
      year={2026},
      eprint={2605.15079},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2605.15079},
}

License

MIT License - see LICENSE file.