GraphWiz: An Instruction-Following Language Model for Graph Computational Problems

Large language models (LLMs) have achieved impressive success across various domains, but their capability in understanding and resolving complex graph problems is less explored. To bridge this gap, we introduce GraphInstruct, a novel instruction-tuning dataset aimed at enabling language models to tackle a broad spectrum of graph problems through explicit reasoning paths. Utilizing GraphInstruct, we build GraphWiz, an open-source language model capable of solving various graph computational problems while generating clear reasoning processes. To further enhance the model's performance and reliability, we integrate the Direct Preference Optimization (DPO) framework within the graph problem-solving context. The improved model, GraphWiz-DPO, achieves an average accuracy of 65% across nine tasks with different complexity levels, surpassing GPT-4 which has an average accuracy of 43.8%. Our study also investigates the relationship between training data volume and model performance, emphasizing the risk of overfitting as data volume increases. Additionally, we explore the transferability of the proposed model across different tasks and datasets, demonstrating its robust zero-shot generalization capability. GraphWiz offers a new blueprint and valuable insights for developing LLMs specialized in graph reasoning and problem-solving.