// the find
llSourcell/Doctor-Dignity
Doctor Dignity is an LLM that can pass the US Medical Licensing Exam. It works offline, it's cross-platform, & your health data stays private.
Doctor Dignity is a fine-tuned Llama2-7B model aimed at offline medical Q&A, targeting people who want health information without sending data to a cloud API. It fine-tunes on medical dialogue data and applies Constitutional AI feedback, then compiles via Apache TVM to run on iOS, Android, and desktop. The README opens with a disclaimer that following its advice could kill you, which is the most honest thing in the repo.
- The offline-first approach is technically sound: TVM compilation to platform-native binaries means actual on-device inference with no server dependency, not just a bundled API key.
- The training pipeline is documented end-to-end — SFT on medical dialogue, then RLCAIF — and the Colab notebook lets someone reproduce the fine-tune without exotic infra.
- 3GB model size is a real engineering constraint that was taken seriously; quantization decisions are explicit rather than just slapping q4 on it and hoping.
- Last commit was September 2023, over two and a half years ago. The MLC LLM toolchain it depends on has changed substantially since then; the build instructions reference nightly wheels and binary blobs that likely no longer resolve correctly.
- Android and Web sections are marked TODO in the README — the cross-platform claim in the description is aspirational, not delivered.
- This is a fork of mlc-llm with a custom fine-tune; the repo contains the full mlc-llm C++/Android/iOS infrastructure (~90% of the files) rather than being a clean composition. Debugging anything means debugging mlc-llm internals.
- The medical fine-tune dataset and RLCAIF reward model details are vague — 'Medical Dialogue Dataset' with no citation, and Constitutional AI here means synthetic feedback from OpenAI, which is an odd dependency for a privacy-first tool.