A pocket medical LLM that warns it could kill you

What it does Doctor Dignity is a 3GB, quantized fine-tune of Meta’s Llama2 7B model built for medical dialogue. The project packages the model for on-device inference via the MLC LLM stack, aiming to keep patient data local. It claims to pass the US Medical Licensing Exam, though the README simultaneously warns that following its advice could cause serious injury or death.

The interesting bit The whole thing is an exercise in shrink-wrapping a specialized LLM for edge deployment: it uses quantization, TVM compilation, and reinforcement learning with constitutional AI to squeeze a domain-tuned model onto a phone. The tension between the USMLE claim and the “do not trust this” disclaimer is the most memorable feature.

Key highlights

• Runs inference locally on iOS (via MLC LLM/TVM), with no API calls or cloud round-trips.
• Weights are available in both quantized-mobile and full-size HuggingFace variants.
• Training pipeline (fine-tuning + RL from constitutional AI feedback) is exposed in a Jupyter notebook.
• Explicitly marked as work-in-progress; Android and Web builds are currently TODO.

Caveats

• The README carries a stark disclaimer: taking the model’s advice seriously could result in injury or death.
• Android and Web support are listed as unfinished; only the iOS build has concrete (if involved) setup documentation.
• The USMLE-passing claim is stated but unsupported by any linked benchmark or evaluation data in the README.

Verdict Worth a look if you are experimenting with on-device quantization and medical LLM fine-tuning, but absolutely not for actual diagnosis or patient interaction. Skip it if you need a production-ready cross-platform health tool.