GraphEval2000
收藏资源简介:
GraphEval2000是由加州大学圣塔芭芭拉分校创建的第一个用于评估大型语言模型(LLMs)在图数据结构问题上的推理能力的数据集。该数据集包含40个图数据结构问题和2000个测试案例,涵盖了稀疏、平面、规则和完整四大类图,每类图又细分为连接、断开、循环和非循环四个子类。数据集的创建过程涉及从LeetCode收集问题,并使用NetworkX生成图样本。GraphEval2000的应用领域主要集中在通过代码挑战评估和提升LLMs在复杂图问题上的推理能力,旨在解决LLMs在处理图结构数据时的局限性。
GraphEval2000 is the first dataset developed by the University of California, Santa Barbara to evaluate the reasoning capabilities of Large Language Models (LLMs) on graph data structure problems. This dataset includes 40 graph data structure problems and 2000 test cases, covering four major graph categories: sparse, planar, regular, and complete. Each category is further subdivided into four subclasses: connected, disconnected, cyclic, and acyclic. The creation of GraphEval2000 involved collecting problems from LeetCode and generating graph samples using the NetworkX library. The primary application scenarios of GraphEval2000 focus on evaluating and enhancing the reasoning abilities of LLMs on complex graph problems via code challenges, aiming to address the limitations of LLMs when processing graph-structured data.

- 1GraphEval2000: Benchmarking and Improving Large Language Models on Graph Datasets加州大学圣塔芭芭拉分校 · 2024年



