GraphGen

Why we included this project

GraphGen tackles a common bottleneck in fine-tuning: getting enough good instruction data. It builds a knowledge graph from your source material and uses it to generate question-answer pairs that fit supervised fine-tuning directly, so you don't have to rely on random samples or hand-write every example. If you're already working with Qwen, Llama, or similar models through LLaMA-Factory or XTuner, the output drops straight into your pipeline. It's particularly useful when you need a lot of domain-specific QA data, and the accompanying paper and cookbook explain the method in detail. For anyone trying to give a smaller model more depth in a specific field, this is a straightforward way to expand the training set.