GRS-QA
收藏资源简介:
GRS-QA是由加州大学圣克鲁兹分校等机构创建的图推理结构化问答数据集,旨在解决现有问答数据集缺乏细粒度推理结构的问题。该数据集包含113,000个基于维基百科的问答对,通过构建推理图来明确捕捉复杂的推理路径。数据集的创建过程包括将每个句子视为节点,并根据原始逻辑关系添加边,同时生成结构负样本以研究结构对问答性能的影响。GRS-QA主要应用于评估大型语言模型在多跳推理任务中的表现,旨在解决复杂推理能力的需求。
GRS-QA is a structured question answering dataset for graph reasoning developed by institutions including the University of California, Santa Cruz, aiming to address the issue that existing QA datasets lack fine-grained reasoning structures. This dataset comprises 113,000 Wikipedia-based question-answer pairs, and explicitly captures complex reasoning paths through the construction of reasoning graphs. The dataset creation workflow involves treating each sentence as a node, adding edges based on original logical relationships, and generating structured negative samples to study the impact of structural features on QA performance. GRS-QA is primarily utilized to evaluate the performance of large language models (LLMs) in multi-hop reasoning tasks, and is designed to meet the demand for assessing models' complex reasoning capabilities.

- 1GRS-QA -- Graph Reasoning-Structured Question Answering Dataset加州大学圣克鲁兹分校 · 2024年



