NJUxlj/Chinese-MedQA-Qwen2

Implements a complete medical QA pipeline from data generation through deployment, featuring custom TRPO trainer alongside SFT/DPO, automated training data synthesis via knowledge graphs (Neo4j, Milvus, LDA), and multi-retrieval algorithms (similarity, BM25, KNN, L2-distance) integrated with FAISS for local knowledge base indexing. The system chains vLLM inference with a multi-agent consultation framework inspired by MDAgents, enabling collaborative diagnosis through tool-calling agents that access medical reference databases and knowledge graphs.