Components¶
Most users only need meddeid to de-identify text. Add another component
when you need to prepare data, review annotations, train a model, or evaluate a
system.
Click a component name to open its GitHub repository. That repository contains the detailed installation instructions, configuration, release notes, and development documentation for the component.
Run de-identification¶
Prepare, review, train, and evaluate¶
These components correspond to optional workflow steps. Use only the ones your project needs.
meddeid-data¶
Use this to import source files into a MedDeID project, create stable document identifiers and splits, validate data, or generate synthetic data.
PyPI: pip install 'meddeid-data[parquet]'
meddeid-annotate¶
Use this local browser application to review identifiers in an assigned JSONL file and save completed primary annotations.
GHCR: docker pull ghcr.io/stighellemans/meddeid-annotate:0.2.0
meddeid-curate¶
Use this only when multiple reviewers annotated the same documents and a curator needs to resolve their differences.
GHCR: docker pull ghcr.io/stighellemans/meddeid-curate:0.2.0
meddeid-subannotate¶
Use this only when a benchmark needs detailed character-level labels inside reviewed identifiers. Ordinary annotation and training do not require it.
GHCR: docker pull ghcr.io/stighellemans/meddeid-subannotate:0.2.0
meddeid-training¶
Use this to train or adapt a model from reviewed data, select and refit a checkpoint, and export a model bundle for inference.
PyPI: pip install 'meddeid-training[train,plots]'
meddeid-eval¶
Use this to score predictions against reviewed test data, compare systems, and create performance and stability analyses.
PyPI: pip install 'meddeid-eval[plots]'
Shared foundations¶
You normally do not choose these packages directly: meddeid and the workflow
tools install the capabilities they need. They are most relevant when you are
integrating with the data contract, developing language behavior, or
contributing to the suite.
meddeid-core¶
Defines the shared record schema, label taxonomy, normalization, stable identifiers, and offset validation used by all components.
PyPI: pip install meddeid-core
meddeid-language-nl¶
Provides Dutch parsing, rendering, resources, and post-processing for nl-BE
and nl-NL.
PyPI: pip install meddeid-language-nl
npm: npm install @meddeid/language-nl@0.2.1
meddeid-language-en¶
Provides English parsing, rendering, resources, and locale-specific
post-processing for en-GB and en-US.
PyPI: pip install meddeid-language-en
npm: npm install @meddeid/language-en@0.2.1
Other useful destinations¶
These are not additional components to install. They are places you may need while trying MedDeID, selecting an artifact, or contributing to a coordinated release.
| Destination | Use it when you want to… |
|---|---|
| MedDeID documentation | Choose a workflow and understand how components work together |
| Non-clinical demo | Try the interface with synthetic or otherwise non-sensitive text |
| Hugging Face collection | Find published models, datasets, benchmark data, and their artifact cards |
meddeid-suite |
Coordinate versions and verify a release across all components |
The suite's publication/ and internal/ directories are maintainer records,
not reader-facing products. External comparison systems are also not MedDeID
components; run them independently and convert their predictions before using
meddeid-eval.
Documentation ownership¶
This site helps you choose a component and follow workflows that cross repository boundaries. Component repositories are authoritative for their API, CLI options, configuration, tests, and release notes. Hugging Face artifact cards are authoritative for model and dataset contents, limitations, licences, and immutable revisions.