stighellemans/meddeid-dutch-synthetic-corpus

MedDeID Dutch synthetic corpus

This repository contains 6,493 synthetic Dutch clinical documents with character-offset de-identification spans. It contains no real patient notes or personal information. The corpus was used to train meddeid-dutch-synth.

Split policy

All 6,493 records are exposed together through one conventional Hugging Face train split. There is no publisher-defined validation split. Here train means the complete model-development corpus; users who tune a model must create and report their own validation partition, including its seed and method.

from datasets import load_dataset

corpus = load_dataset(
    "stighellemans/meddeid-dutch-synthetic-corpus",
    split="train",
)
parts = corpus.train_test_split(test_size=0.10, seed=42)

Record contract

Each JSONL row uses the canonical MedDeID document schema:

• —document_id: stable synthetic document identifier;
• —text: source document text;
• —spans: primary PII spans with absolute, half-open Unicode-codepoint offsets;
• —metadata_json: available synthetic-generation metadata encoded as JSON;
• —optional workflow fields retained from the source artifact.

spans is the only accepted top-level span container. Each span uses begin, end, text, label, category, and subtype. Unknown extra fields are preserved as recorded rather than rewritten for publication.

The viewer file uses metadata_json because the source metadata contains many valid but heterogeneous nested structures that cannot be represented by one automatically inferred Arrow struct. The unchanged authoritative JSONL remains available as source/original.jsonl; decoding metadata_json reconstructs the same metadata object for every row.

Annotation guideline

The current English and Dutch ProductionLabels_v1 guidelines are included in guidelines/ as PDF and DOCX files.

Integrity

The release-level SHA-256 file is included as CHECKSUMS.sha256. It covers both the viewer file and the unchanged source payload.

Creation and privacy

The documents were generated from synthetic clinical cases and synthetic Belgian identifiers. No real patient note or real patient identifier was provided during generation. The corpus is useful for de-identification model development, but synthetic language does not reproduce every property of real clinical documentation.

Project and institutional context

Developed by Stig Hellemans, Tom Stroobants, Elyne Scheurwegs, Pieter Meysman, Philippe Jorens and Kris Laukens at the University of Antwerp and Antwerp University Hospital (UZA). This work is supported by a Research Foundation Flanders (FWO) SB Fellowship, grant 1SA3226N.

Author ORCIDs: Stig Hellemans, Tom Stroobants, Elyne Scheurwegs, Pieter Meysman, Philippe Jorens, and Kris Laukens.

Citation

@dataset{hellemans_2026_meddeid_synthetic_corpus,
  author    = {Hellemans, Stig and Stroobants, Tom and Scheurwegs, Elyne and Meysman, Pieter and Jorens, Philippe and Laukens, Kris},
  title     = {MedDeID Dutch synthetic corpus},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/stighellemans/meddeid-dutch-synthetic-corpus}
}

@misc{hellemans2026meddeid,
  author        = {Hellemans, Stig and Stroobants, Tom and Scheurwegs, Elyne and Meysman, Pieter and Jorens, Philippe G. and Laukens, Kris},
  title         = {{MedDeID} enables locally governed clinical-text de-identification from real or synthetic training data},
  year          = {2026},
  eprint        = {2609.10049},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2609.10049}
}

When reporting experiments, also record the immutable dataset revision and the validation-split method and seed.

Licence

The dataset and included annotation guidelines are licensed under the Creative Commons Attribution 4.0 International licence (CC BY 4.0). Cite the MedDeID dataset when using this resource.