A 4-billion-parameter vision-and-language model specialising in GLP-1–related diabetes care, obtained by merging our LoRA adapters into Google’s MedGemma-4B-IT base model.
It answers text-only queries (image grounding unchanged).

🩺 Why we built this model

GLP-1 agonists (semaglutide, liraglutide, etc.) are rapidly transforming Type-2 diabetes care.
Existing open LLMs lack up-to-date, peer-reviewed knowledge and struggle with the empathetic tone required in doctor-patient dialogue.
Our goal is clinical decision support and patient education, not automated diagnosis.

1 Data preparation 🗄️

Layer	Sources	Rationale
1️⃣ Medical guidelines & protocols	~100 recent (≤ 3 y) U.S. endocrine society documents	Guarantees latest treatment algorithms
2️⃣ Scientific studies & reviews	~200 peer-review papers / meta-analyses on semaglutide & peers	High-evidence risk/benefit data
3️⃣ General medical knowledge	Textbooks + open datasets (e.g. MedMCQA)	Terminology & pathophysiology grounding
4️⃣ Real-world consultations (anonymised; consented)	Transcripts of doctor–patient visits	Teaches natural, empathetic tone (↑ satisfaction ×1.5–2; Johri 2023)
5️⃣ Safety / awareness corpora	Adversarial & sensitive-topic prompts	Model defers to clinicians when unsure

All 100 k QA/dialogue records were de-duplicated, cleaned, and medically reviewed (Clusmann 2023).

2 Training pipeline ⚙️

Phase	Method	Goal
Base	`google/medgemma-4b-it` FP16	Vision-language backbone
SFT-LoRA	Supervised fine-tuning on GLP-1 QA	Teach domain facts
DPO-LoRA	Direct Preference Optimisation (Dubey 2024) on paired outputs	Align tone & user preference (concise + empathetic)
Merge	`merge_and_unload()` → 3-shard .safetensors	Single checkpoint for inference

Hyper-params (SFT): LR 5e-5 · batch 64 (grad-acc 16 × 2 GPU) · cosine schedule · 2 epochs
DPO used the same hardware with a 0.1 preference LR.

3 Intended use ✅

Patient education on GLP-1 therapy, lifestyle, side-effects.
Clinical decision support tools with a licensed professional in the loop.
Research on safe medical-AI interaction patterns.

4 Limitations ⚠️

Not a medical device. No autonomous diagnosis, prescription or emergency triage.
Knowledge cut-off ≈ Nov 2024 (last guideline snapshot).
English-centric; performance on other languages untested.
Does not improve image grounding—the original MedGemma capabilities apply.

5 Safety & ethical compliance 🔒

We follow the Reforma Health SLM validation & ethics framework:

Data governance – HIPAA-compatible storage, PHI stripping, dual expert audit.
Clinical QA – 2-tier review (physician + pharm.D) for every model change.
Bias testing – evaluate across age, sex, ethnicity, insurance status.
Harm safeguard – refusal / hand-off when confidence < 0.65 or question out-of-scope.
Continual monitoring – nightly hallucination & toxicity sweeps; flagged outputs re-enter DPO set.

6 Quick start 🔧

python

from transformers import AutoModelForImageTextToText, AutoProcessor
import torch

model_id = "YmaHealth/medgemma4b_yma_health_merged"

model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,   #  use float16 on GPUs without bfloat16
    device_map="auto",
)
processor = AutoProcessor.from_pretrained(model_id)

messages = [
    {
        "role": "system",
        "content": [{"type": "text", "text": "You are an endocrinologist specialised in GLP-1 therapy."}]
    },
    {
        "role": "user",
        "content": [{"type": "text", "text": "How should I titrate semaglutide if nausea persists?"}]
    }
]

inputs = processor(messages, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0], skip_special_tokens=True))

YmaHealth/medgemma_4b_it_glp

YmaHealth · MedGemma-4B-IT GLP-1 Specialist