Our Open Source Position
Open where it should be. Closed where it must be.
A clean legal and philosophical line, drawn in the same place every time.
We open-source
Training recipes and pipelines
The scripts, configurations, and dataflows we use to produce healthcare-tuned models. Reproducible builds, not black boxes.
Evaluation benchmarks
Task-specific evaluation suites for healthcare workflows — the credibility backbone of tuning-as-a-service. Anyone can run them. Anyone can propose better ones.
Synthetic & de-identified datasets
Training and evaluation data derived from public sources, safely de-identified, or generated from scratch. Enough for anyone to reproduce our published results.
Deployment tooling
The platform code that runs inference, audit logging, and access control on the hospital server. Auditable by any security team.
Base model weights on public data
Our starting-point tuned models, trained only on public and synthetic data. Free for anyone to download, evaluate, or build on.
We never open-source
Anything trained on real patient data
Weights that ever touched PHI stay inside the walls where that data lives. No exceptions, no exceptions with anonymization, no exceptions with paper agreements. If it was trained on your data, it belongs to you.
Client-specific adapters
The LoRA adapters we train for a specific hospital’s workflows are the client’s property. They stay on the client’s hardware. GofarAI does not retain copies.
Client evaluation results
How your specific tuning run performed against your specific data is your business. We publish patterns and aggregate improvements, never client-identifying results.
Anything that could reidentify
Any artifact where reidentification is even theoretically possible stays private. “Probably safe” is not a standard we accept for PHI.
Weights that ever touched PHI
Once weights have been trained on real patient data, they stay in the environment where that data lives. The bright line is at the hospital’s perimeter, not at the model file.
Why this line, in this place.
Open source is a moat, not a marketing checkbox. When training recipes, evaluation suites, and base weights are public, hospitals can verify what they’re getting. Compliance teams can audit our claims. The ML community can push back on our methods.
But PHI is different. Real patient data isn’t shareable, no matter how the model is described. Drawing the line here — open recipes, closed patient-trained weights — is what makes both halves of this business defensible.
Read the recipes. Run the evaluations.
Public repositories, model cards, and evaluation benchmarks are rolling out with our first releases. Talk to us if you want early access or want to contribute.