vaibhavrakshe161/medqa-hindi

BERTScore | 0.71 | 0.82 | 0.87 | ↑22% | | Medical Accuracy | 45% | 72% | 78% | ↑73% |

Evaluation Methodology

• —ROUGE-L: Measures longest common subsequence between generated and reference answers
• —BERTScore: Semantic similarity using multilingual BERT embeddings
• —Medical Accuracy: Keyword overlap on medical entities (medications, symptoms, body parts, treatments)

📁 Project Structure

medqa-hindi/
├── data/
│   ├── prepare_dataset.py      # Dataset download & formatting
│   ├── translation.py          # IndicTrans2 EN→HI translation
│   ├── create_dpo_pairs.py     # Self-play DPO pair generation
│   └── processed/              # Generated datasets (gitignored)
├── training/
│   ├── qlora_finetune.py       # QLoRA training script
│   ├── dpo_align.py            # DPO alignment script
│   └── configs/
│       ├── qlora.yaml          # QLoRA hyperparameters
│       └── dpo.yaml            # DPO hyperparameters
├── evaluation/
│   ├── evaluate.py             # Metrics computation
│   ├── compare_models.py       # Report generation
│   └── results/                # Evaluation outputs (gitignored)
├── api/
│   ├── main.py                 # FastAPI application
│   ├── inference.py            # Model loading & generation
│   └── models.py               # Pydantic schemas
├── frontend/
│   ├── index.html              # Single-page application
│   ├── styles.css              # Dark medical theme
│   ├── app.js                  # Frontend logic
│   └── charts.js               # Chart.js visualizations
├── requirements.txt
└── README.md

💻 Hardware Requirements

Training (Kaggle T4)

Inference

⚠️ License & Disclaimer

Medical Disclaimer

This system is for educational purposes only. It is not 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.

यह जानकारी केवल शैक्षिक उद्देश्यों के लिए है। यह पेशेवर चिकित्सा सलाह का विकल्प नहीं है। किसी भी स्वास्थ्य संबंधी निर्णय से पहले अपने चिकित्सक से परामर्श करें।

License

MIT License - See LICENSE file for details.

The model weights and datasets may have their own licenses:

• —Qwen2.5: Tongyi Qianwen License
• —MedQuAD: Public Domain
• —MedMCQA: Apache 2.0
• —IndicTrans2: MIT

🙏 Acknowledgments

• —Qwen Team for the base model
• —HuggingFace for datasets and transformers
• —AI4Bharat for IndicTrans2
• —Kaggle for free GPU access

📬 Contact

For questions or contributions, please open an issue on GitHub.

Built with ❤️ for better healthcare AI in Indian languages.