An AI-powered health question-answering system that supports both Bengali and English queries.
Built on top of Llama-3.2-3B-Instruct (Unsloth) with LoRA fine-tuning, this project enables multilingual healthcare assistance with optimized performance and API deployment.

• Model: unsloth/Llama-3.2-3B-Instruct
• LoRA Fine-tuning: Efficient training with low-rank adaptation
• Bilingual Support: Automatically detects question language (Groq API + Deepseek model)
• Dataset Sources:
  • Medical Customer Care (English)
  • Doctor QA Bangla
• Performance: Achieved validation loss: 0.74
• Deployment: FastAPI + ngrok, tested with Postman
• Environment: Trained on Kaggle GPU

• Language Detection: Groq API + Deepseek

• English Generation:

max_new_tokens=128, temperature=0.1, min_p=0.1

• Bengali Generation:

max_new_tokens=256, temperature=0.1, min_p=0.1

• Repetition handled via unique sentence filtering.

• Built with FastAPI
• Exposed via ngrok
• Tested with Postman

✅ Bilingual support (Bengali + English)
✅ Validation loss: 0.74
✅ Stable inference with low temperature
✅ Deployed as REST API