GraphInstruct

该数据集名为GraphInstruct，包含了九种不同时间复杂度的图推理问题，涵盖了线性、多项式以及NP完全任务。该数据集旨在评估模型在处理多项式时间任务时的性能，展示了模型处理不断增长问题复杂性的能力。数据集中的图包含了最多1000个节点。任务类型包括循环检测、连通性、二分图检查、拓扑排序、最短路径、最大三角形和以及最大流等多种图推理问题。This dataset, named GraphInstruct, encompasses graph reasoning problems with nine distinct time complexities, covering linear, polynomial, and NP-complete tasks. It is developed to evaluate models' performance in handling polynomial-time tasks and showcase their ability to cope with escalating problem complexity. Each graph in this dataset contains up to 1000 nodes. The included task types span various graph reasoning problems, including cycle detection, connectivity, bipartite graph checking, topological sorting, shortest path finding, maximum triangle sum, and maximum flow.

搜集汇总

背景与挑战

背景概述

GraphInstruct是一个用于增强大型语言模型图理解和推理能力的基准数据集，基于论文《GraphInstruct: Empowering Large Language Models with Graph Understanding and Reasoning Capability》提出。该数据集是动态生成的，支持从零开始生成和评估，并提供了详细的生成、评估和模型训练步骤，包括环境安装、脚本运行和基于LLaMAFactory的监督微调。