MedGemma-4B Cardiology (GGUF)
uaritm — Domain-Finetuned Clinical LLM for Cardiology Based on Google MedGemma-4B-IT
Model Description
MedGemma-4B-Cardiology is a domain-adapted medical language model built on
google/medgemma-4b-it and fine-tuned via supervised instruction tuning (SFT) on anonymized cardiology clinical data:
- discharge summaries
- history of illness
- surgical procedures
- treatment pathways
- physician recommendations
The model is optimized for:
- generating clinically grounded cardiology recommendations
- structuring and summarizing medical documentation
- rewriting clinical narratives
- assisting in cardiology decision support
GGUF format ensures compatibility with:
- llama.cpp
- Ollama
- LM Studio
- GPT4All
- koboldcpp
Architecture
- Base model: Google MedGemma-4B-IT
- Parameters: ~4B
- Fine-tuning: LoRA + SFT
- Precision: Q4_K_M / Q5_K_M GGUF
- Context window: up to 4096 tokens
Repository Contents
medgemma-4b-cardiology.Q4_K_M.gguf
medgemma-4b-cardiology.Q5_K_M.gguf
tokenizer.json
README.md
Usage Examples
llama.cpp
./main -m medgemma-4b-cardiology.Q4_K_M.gguf
-p "You are a cardiologist. Based on the clinical details, generate treatment recommendations."
Ollama
Create a file named Modelfile:
FROM ./medgemma-4b-cardiology.Q4_K_M.gguf
TEMPLATE "<start_of_turn>user\n{{ .Prompt }}\n<end_of_turn>\n<start_of_turn>assistant"
PARAMETER temperature 0.3
Then run:
ollama create medcardio -f Modelfile
ollama run medcardio
Python (llama-cpp-python)
from llama_cpp import Llama
llm = Llama(
model_path="medgemma-4b-cardiology.Q4_K_M.gguf",
n_ctx=4096,
n_gpu_layers=35,
)
prompt = "You are a cardiologist. Patient: 67-year-old male with stable angina. Provide evidence-based recommendations."
output = llm(prompt, max_tokens=400)
print(output["choices"][0]["text"])
Prompt Format (MedGemma Chat Template)
The model expects chat-structured input:
<start_of_turn>user
[Your instruction or clinical text]
<end_of_turn>
<start_of_turn>assistant
Example:
<start_of_turn>user
You are a cardiologist. Generate recommendations for:
Sex: Male
Age: 67
Length of stay: 10 days
[discharge summary...]
<end_of_turn>
<start_of_turn>assistant
Fine-Tuning Details
Training dataset included >12,000 cardiology-specific instruction-response pairs:
{
"prompt": "...",
"response": "..."
}
Converted into Gemma-style messages:
{
"messages": [
{"role": "user", "content": "..."},
{"role": "assistant", "content": "..."}
]
}
Training setup:
- LoRA r=16, alpha=32
- BF16 training
- max_length: 2048–4096
- learning rate sweep: 5e-5 → 3e-4
- warmup_ratio: 0.03–0.10
- gradient checkpointing: enabled
- optimizer and scheduler from TRL SFTTrainer
Disclaimer
This model is NOT a medical device and must not be used for autonomous diagnosis or treatment decisions.
Outputs require interpretation and verification by certified medical professionals.
Intended solely for research and decision-support augmentation.
Limitations
- Optimized for cardiology and cardiac surgery
- Reduced accuracy outside these domains
- No vision capabilities (text-only MedGemma IT)
- May generate incomplete or generalized recommendations
Citing & Authors
If you use this model in your research, please cite:
@misc{Ostashko2025MedGemmaCardiology,
title = {MedGemma-4B-Cardiology: A Domain-Finetuned Clinical LLM for Cardiology},
author = {Uaritm},
year = {2025},
url = {ai.esemi.org}
}
HuggingFace
If you found this model useful — please give a star on HuggingFace:
