InternScience/GraphGen

Constructs fine-grained knowledge graphs from source documents, then uses expected calibration error metrics to identify knowledge gaps in LLMs and prioritize high-value QA generation with multi-hop neighborhood sampling and style-controlled diversity. Integrates with Ray for distributed pipeline execution, supports multiple LLM backends (vLLM, HuggingFace Transformers, SGLang, Ollama), and provides generated data compatible with LLaMA-Factory and xtuner for downstream fine-tuning workflows.

978 stars. Actively maintained with 9 commits in the last 30 days.