Medical Gemma 3 270M — Fine-Tuned
A small language model fine-tuned on medical Q&A data to act as a domain-specific medical assistant.
Model Details
- Base Model : google/gemma-3-270m-it
- Dataset : ChatDoctor-HealthCareMagic (4,500 samples)
- Training : 2 epochs, RTX 3050 (4.3GB VRAM)
- Framework : HuggingFace Transformers + TRL SFTTrainer
What it Does
Answers medical questions in a conversational doctor-patient style learned from real doctor responses.
Example
Input: What are the symptoms of diabetes?
Output: Hi, Thanks for asking. Diabetes is a condition in which the body has an excess of sugar. It is usually caused by high blood sugar levels...
Metrics vs Base Model
| Metric | Base Model | Fine-Tuned |
|---|---|---|
| Avg ROUGE-1 | 0.2577 | 0.1724 |
| Avg ROUGE-2 | 0.1139 | 0.0593 |
| Avg ROUGE-L | 0.2082 | 0.1356 |
| Avg Response Time | 6.10s | 3.01s |
| Avg Response Length | 104.8 | 63.8 |
Key Observations
- Fine-tuned model is 2x faster than base model
- Learned concise doctor-style conversational responses
- Lower ROUGE scores reflect focused answers not longer generic responses
Disclaimer
This model is for educational purposes only. Not intended for real medical advice or diagnosis. Always consult a qualified doctor.
Author
Divyansh Vats — github.com/vatsdivyansh
- Downloads last month
- 38