MedDroid-v9

A style-tuned fine-tune of Google's MedGemma 4B for the MedDroid AI medical assistant (medicalandroid.com). It keeps MedGemma's medical knowledge and image-reading, and adds a warmer, more structured, multilingual (English / தமிழ் / हिन्दी) voice.

⚠️ Educational use only — NOT a medical device and NOT for clinical decision-making. MedDroid-v9 provides general health information, not a diagnosis or a prescription. It can be wrong. Always consult a qualified doctor or pharmacist. In an emergency, contact your local emergency number.

---

What it is

MedGemma 4B is already strong at medical text and medical imaging. Our benchmark showed its real gaps versus a frontier model were language and presentation — not knowledge:

• It answered Indian-language questions in English
• Answers were terse and inconsistently structured
• Safety framing ("not a diagnosis", red-flags) was hit-or-miss

Earlier attempts to fix this by fine-tuning on medical Q&A corrupted the model's drug-dosing facts (a known risk of SFT on small models). So MedDroid-v9 was trained on language + tone/format examples ONLY, with zero dose numbers or hard facts in the data — there is nothing factual for the fine-tune to overwrite. The vision tower is frozen.

What improved (vs base MedGemma)

Dimension	Base MedGemma	MedDroid-v9
Tamil / Hindi questions	answered in English	answers in the same language ✅
Structure & depth	terse	headings, bullets, "when to see a doctor" ✅
Safety framing	inconsistent	consistent "not a diagnosis" + red-flags ✅
Chest X-ray (COPD)	correct	still correct (hyperinflation → COPD) ✅
Paracetamol dose	correct (4 g)	still correct (4 g) ✅
Dangerous errors on our benchmark	0	0 ✅

⚠️ Known limitations

• Exact drug dosing can be imprecise. Common doses (e.g. paracetamol) are reliable, but the model may be imprecise for specific drugs (e.g. it confused amoxicillin with a co-amoxiclav figure). Do not rely on it for exact dosing — MedDroid pairs it with retrieval/guardrails and a frontier-model fallback for dosing in production.
• It is a 4B model: less capable than frontier models on long, complex reasoning.
• It can hallucinate and occasionally appends slightly off boilerplate. Treat every output as general information to verify with a clinician.
• Not evaluated for any regulated clinical purpose; no ISO/CLIA/CDSCO/FDA validation.

Intended use

• Consumer health education: explaining symptoms, medicines, lab reports and scans in plain language, in English/Tamil/Hindi.
• A component in a compound system (with retrieval for facts and a frontier-model fallback), not a standalone clinical tool.

Out of scope: diagnosis, treatment decisions, prescribing, emergency triage as a sole source, or any clinical/regulated use.

How to use

This repo ships GGUF files — the language model and the vision projector (mmproj) for image input.

Ollama

Create a Modelfile next to the two .gguf files (both FROM lines enable vision):

FROM ./medgemma-4b-it.Q4_K_M.gguf
FROM ./medgemma-4b-it.F16-mmproj.gguf
PARAMETER temperature 0.6
PARAMETER num_ctx 4096
SYSTEM """You are MedDroid, an AI medical assistant. Give clear, genuinely useful general information. Be warm and concise. You are not the treating clinician: no definitive diagnosis or individualised prescription. Detect the user's language and reply in it. Lead with emergency advice for red-flag symptoms."""

ollama create meddroid-v9 -f Modelfile
ollama run meddroid-v9 "எனக்கு காய்ச்சல் மற்றும் தொண்டை வலி. என்ன செய்ய வேண்டும்?"

llama.cpp (with vision)

llama-mtmd-cli -m medgemma-4b-it.Q4_K_M.gguf --mmproj medgemma-4b-it.F16-mmproj.gguf

Training

• Method: QLoRA (Unsloth), 4-bit base, LoRA on language layers only; vision tower frozen.
• Data: ~24 hand-written language + tone/format examples (English/Tamil/Hindi), weighted; no dose numbers or factual medical claims (by design).
• Hyperparameters: rank 8, alpha 16, dropout 0.05, LR 2e-5, 3 epochs, cosine schedule, seq len 2048.
• Hardware: free Kaggle T4 ×2.
• Data policy: open/own data only. No outputs of other proprietary models were used for training.

License & attribution

MedDroid-v9 is a derivative of google/medgemma-4b-it and is distributed under the Health AI Developer Foundations (HAI-DEF) terms — see the license link above. By using this model you agree to those terms, including the prohibited-use policy. This model is not affiliated with or endorsed by Google or Anthropic.

Citation

Author: Dr. Sanjay Anbu — MedDroid (medicalandroid.com) Built on Google MedGemma 4B. Educational project; not a certified medical device.