Gemma-2-2B Medical Q&A Assistant 🏥

This is a fine-tuned version of google/gemma-2-2b-it specialized for medical question answering. The model has been trained on real patient-doctor conversations to provide helpful, accurate medical information with appropriate safety disclaimers.

🎯 Model Details

• Base Model: Gemma-2-2B-Instruct (2B parameters)
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Quantization: 4-bit (nf4) for efficient training and inference
• Dataset: ChatDoctor-HealthCareMagic (1,500 curated samples)
• Domain: Medical Q&A, Healthcare Information
• Use Case: Medical information assistant, health education, preliminary medical guidance

🔧 Training Configuration

LoRA Parameters

• Rank (r): 8
• Alpha: 16
• Dropout: 0.1
• Target Modules: All linear layers

Training Hyperparameters

• Epochs: 3
• Learning Rate: 2e-5
• Batch Size: 4 per device
• Max Sequence Length: 1024 tokens (for detailed medical responses)
• Scheduler: Cosine with 10% warmup
• Optimizer: AdamW with weight decay (0.01)
• Precision: BF16
• Gradient Clipping: 1.0

Training Infrastructure

• Hardware: Single GPU (T4/V100/A100)
• Training Time: ~30-45 minutes
• Framework: TRL + Transformers + PEFT

📊 Dataset

The model was fine-tuned on a curated subset of the ChatDoctor-HealthCareMagic-100k dataset, which contains:

• Real patient questions from HealthCareMagic platform
• Professional doctor responses
• Diverse medical topics (symptoms, treatments, medications, preventive care)
• Privacy-preserved and ethically sourced data

Training Samples: 1,500 high-quality medical Q&A pairs

💻 Usage

Basic Inference

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "sweatSmile/Gemma-2-2B-MedicalQA-Assistant"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Format your medical query
prompt = """<start_of_turn>user
I've been experiencing persistent headaches for the past week. What could be causing this?<end_of_turn>
<start_of_turn>model
"""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.9,
    do_sample=True
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

4-bit Quantized Inference (Recommended)

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16
)

model = AutoModelForCausalLM.from_pretrained(
    "sweatSmile/Gemma-2-2B-MedicalQA-Assistant",
    quantization_config=quantization_config,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("sweatSmile/Gemma-2-2B-MedicalQA-Assistant")

🩺 Example Outputs

Query: "What are the common symptoms of diabetes?"

Response:

Common symptoms of Type 2 diabetes include frequent urination, increased thirst, unexplained weight loss, fatigue, blurred vision, slow-healing wounds, and tingling or numbness in hands or feet. If you're experiencing multiple symptoms, it's important to get your blood sugar levels checked. Early detection and management are crucial.

⚕️ Disclaimer: This information is for educational purposes only. Please consult a qualified healthcare professional for medical advice.

---

Query: "How can I manage high blood pressure naturally?"

Response:

Natural management of high blood pressure includes: regular exercise (30 minutes daily), reducing sodium intake, maintaining healthy weight, limiting alcohol, managing stress through meditation or yoga, eating potassium-rich foods (bananas, leafy greens), and ensuring adequate sleep. However, these should complement, not replace, prescribed medications.

⚕️ Disclaimer: This information is for educational purposes only. Please consult a qualified healthcare professional for medical advice.

⚠️ Important Limitations & Safety

Medical Disclaimers

• 🚨 NOT A REPLACEMENT FOR PROFESSIONAL MEDICAL ADVICE: This model provides general health information only
• 🚨 Always consult a qualified healthcare provider for diagnosis and treatment
• 🚨 Emergency situations: Call emergency services immediately for urgent medical conditions
• 🚨 Individual variation: Medical advice must be personalized to your specific situation

Model Limitations

• Trained on 1,500 samples - may not cover all medical scenarios
• Limited to English language medical terminology
• May not reflect the most recent medical research (knowledge cutoff dependent on base model)
• Cannot perform physical examinations or order diagnostic tests
• May occasionally generate plausible-sounding but incorrect information (hallucinations)
• Not validated against clinical benchmarks

Appropriate Use Cases ✅

• General health education and information
• Understanding common medical terms and conditions
• Preliminary research before doctor's appointments
• Health literacy improvement
• Medical training assistance (with supervision)

Inappropriate Use Cases ❌

• Self-diagnosis or self-treatment
• Emergency medical situations
• Replacing professional medical consultations
• Making critical healthcare decisions
• Prescribing medications or treatments
• Mental health crisis intervention

🎓 Intended Use

This model is designed for:

• Educational purposes: Teaching and learning about common health conditions
• Health information access: Providing accessible medical knowledge
• Research: Medical AI and NLP research
• Prototyping: Building healthcare chatbot prototypes
• Medical training: Supplementary tool for medical students (with instructor oversight)

📈 Performance Notes

• Strengths: Common conditions, preventive care, general wellness advice, medical terminology
• Best performance on: Questions similar to training distribution (patient-doctor Q&A format)
• Quantization: 4-bit quantization maintains ~95% of full precision performance with significant memory savings

🔬 Technical Specifications

📝 Citation

@misc{gemma2-medical-qa-2025,
  author = {sweatSmile},
  title = {Gemma-2-2B Medical Q&A Assistant},
  year = {2025},
  publisher = {HuggingFace},
  journal = {HuggingFace Model Hub},
  howpublished = {\url{https://huggingface.co/sweatSmile/Gemma-2-2B-MedicalQA-Assistant}}
}

🙏 Acknowledgments

• Base Model: Google's Gemma team for the excellent Gemma-2-2B-Instruct model
• Dataset: ChatDoctor team and HealthCareMagic for the medical Q&A dataset
• Framework: HuggingFace TRL, Transformers, and PEFT libraries

📜 License

This model inherits the Gemma License from the base model. The fine-tuned weights are released under the same terms.

Usage Restrictions: Please review Google's Gemma Terms of Use, particularly regarding healthcare applications.

⚖️ Ethical Considerations

• Model outputs should always include medical disclaimers
• Designed to encourage users to seek professional medical advice
• Training data sourced ethically from publicly available patient-doctor interactions
• No personally identifiable information (PII) in training data
• Built with safety alignment from base Gemma-2 model

🔄 Version History

• v1.0 (Current): Initial release with 1.5k samples, LoRA fine-tuning

---

Disclaimer: This model is a research prototype and educational tool. It is not intended for clinical use or as a substitute for professional medical advice, diagnosis, or treatment. Always seek the advice of your physician or other qualified health provider with any questions you may have regarding a medical condition.