jiangxinke/Agentic-RAG-R1

Implements GRPO (Generalized Relevance Policy Optimization) to train language models with autonomous tool-calling and multi-step reasoning over retrieval actions, supporting an agent memory stack with backtracking and summarization. Integrates with ArtSearch for Wikipedia retrieval and TCRAG as a rollout generator, while offering LoRA tuning, quantization, and DeepSpeed distributed training (Zero 2/3) to efficiently fine-tune models up to 32B on 2 A100 GPUs. Includes a composite reward model combining accuracy, format, and RAG-specific RAGAS-based scoring for optimizing both answer quality and retrieval effectiveness.

393 stars.