Public artifacts¶
The MedDeID Hugging Face collection contains the current default model and datasets. Each item has its own page with usage guidance, limitations, licence, and version history.
Model¶
stighellemans/meddeid-dutch-synth¶
The current Dutch clinical de-identification model, trained on synthetic data. Download and run the complete model through the MedDeID package:
For a reproducible or air-gapped deployment, save the exact model version and keep its files together.
Datasets¶
stighellemans/meddeid-dutch-synthetic-corpus¶
The complete 6,493-document synthetic dataset used for model development. Divide this dataset into training and validation data; keep the independent benchmark below separate.
stighellemans/meddeid-dutch-synthetic-benchmark¶
An independent set of 300 synthetic documents for final testing. It includes extra reviewed detail for measuring whether the important parts of each identifier were removed.
Annotation guidelines¶
The current ProductionLabels_v1 guidelines are available in English and Dutch. PDF reading copies and editable DOCX sources are included in the Zenodo v2 release. They explain what annotators should mark; the precise software rules are documented separately in meddeid-core.
Artifact documentation belongs with the artifact¶
Model and dataset cards are authoritative for:
- exact files, row counts, and checksums;
- intended use and out-of-scope use;
- training or generation provenance;
- evaluation results and limitations;
- licences and required attribution;
- immutable revisions and citations.
This site explains how artifacts fit into suite workflows. It should not copy full cards, because copied limitations and revisions become stale.
Public demo¶
The MedDeID hosted demo
runs the public synthetic model for non-sensitive examples. It executes on
Hugging Face infrastructure: never paste real patient or caregiver information.
Use meddeid locally for clinical text.