Efficient Supervised Fine Tuning for LLM-Based Graph Data Simulation

This paper presents a novel framework for controllable graph generation through supervised fine-tuning with dynamic validation mechanisms. We propose dual prompting strategies combining rule-driven specifications and statistical-aware constraints, integrated with parameter-efficient LoRA adaptation, to generate both algorithmically defined graphs (e.g., ER and BA models) and real-world networks (social/biological graphs). Experimental results demonstrate that our stepwise node-by-node validation approach significantly improves topological accuracy while maintaining the generalization capabilities of pretrained language models, offering a practical solution for high-fidelity graph synthesis.