MatchMiner is built in the open so other centers can use it, inspect how matches are made, and run it without sending patient data anywhere. The code, models, and data below are published by Dana-Farber.
Code
All published MatchMiner code lives in the dfci GitHub organization.
| Repository | What it is | License |
|---|---|---|
| matchminer-ai-inference | MatchMiner-AI inference pipeline, published as the matchminer-ai Python package. Documentation. |
CC BY-NC-ND 4.0 |
| matchminer-ai-training | Code for training the MatchMiner-AI models | CC BY-NC-ND 4.0 |
| matchengine-V2 | MatchMiner Genomics matching engine | Apache-2.0 |
| matchminer-api | MatchMiner Genomics API, which stores patient data and match results | See repository |
| matchminer-ui | MatchMiner Genomics web interface | AGPL-3.0 |
The matchminer repository is the starting point for a MatchMiner Genomics deployment. MatchMiner Dashboard’s code will be released in winter 2026–27.
Models and data
MatchMiner-AI’s models and training data are on Hugging Face under ksg-dfci (Dana-Farber’s Knowledge Systems Group). The models are trained entirely on synthetic medical records, which is why they can be shared.
Model families
MatchMiner-AI also uses an open-source large language model, run with the prompts and parameters in the inference pipeline. The models the team trained fall into three families:
- TrialSpace finds candidate trials for a patient, for example TrialSpace-0526
- TrialChecker judges whether the patient meets a trial’s eligibility criteria, for example TrialChecker-0526
- BoilerplateChecker checks the common exclusion criteria most trials share, for example BoilerplateChecker-0526
The current releases are MatchMiner-AI 0426 and 0526. ONNX versions of the 0526 models, a portable format that runs without PyTorch, are also available.
Synthetic training data
- mmai-synthetic-0526, about 15.5 million rows
- Earlier releases: mmai-synthetic-0426 and mmai-synthetic-1225
Who else uses it
See Who uses MatchMiner for cancer centers that run MatchMiner or build on it.
Contributing
Bug reports and pull requests are welcome in each repository. For larger collaborations, contact the team.