GraCoRe
收藏资源简介:
GraCoRe数据集由哈尔滨工业大学(深圳)和鹏城实验室共同创建,旨在评估大型语言模型在图理解和复杂推理方面的能力。该数据集包含5140个图,涵盖纯图和异构图,通过19个任务测试模型的10种不同能力。数据集的创建过程精细,通过调整图的大小和网络稀疏度来控制复杂性。GraCoRe主要应用于社交网络分析、药物发现等领域,旨在解决图结构数据理解和推理的问题。
The GraCoRe dataset was co-developed by Harbin Institute of Technology (Shenzhen) and Peng Cheng Laboratory, aiming to evaluate the capabilities of large language models (LLMs) in graph understanding and complex reasoning. It comprises 5,140 graphs, including both homogeneous and heterogeneous graphs, and assesses 10 distinct capabilities of models across 19 tasks. The dataset was constructed with meticulous procedures, where complexity is controlled by adjusting graph sizes and network sparsity. GraCoRe is primarily utilized in domains such as social network analysis and drug discovery, and is designed to tackle challenges related to graph-structured data understanding and reasoning.
GraCoRe
摘要
GraCoRe 是一个用于系统评估大型语言模型(LLMs)在图理解与复杂推理能力的基准测试。该基准测试通过一个三层分层分类法,对纯图和异构图进行分类和测试,细分为10个不同的能力领域,并通过19个任务进行测试。GraCoRe 包含11个数据集,共5,140个不同复杂度的图。在评估中,使用了三个闭源和七个开源的LLMs,并从能力和任务的角度进行了全面分析。主要发现包括:语义增强提高了推理性能,节点顺序影响任务成功,处理较长文本的能力并不一定能提高图理解或推理能力。

- 1GraCoRe: Benchmarking Graph Comprehension and Complex Reasoning in Large Language Models哈尔滨工业大学(深圳),鹏城实验室 · 2024年



