HRMCR (HAE-RAE Multi-Step Commonsense Reasoning)
收藏资源简介:
HRMCR数据集是由延世大学的研究团队创建的一个多步推理基准测试,旨在评估大语言模型在韩国文化背景下的推理能力。该数据集包含两个子集:Date和Zodiac,每个子集包含50个问题,总计100个问题。数据集通过模板和算法自动生成,要求模型在推理过程中整合韩国文化知识。数据集的应用领域主要集中在大语言模型的推理能力评估,特别是针对韩国文化和语言的多步推理任务。通过这一数据集,研究者可以更好地理解模型在复杂推理任务中的表现,并探索模型在文化特定背景下的推理能力。
The HRMCR dataset is a multi-step reasoning benchmark developed by a research team from Yonsei University, which aims to evaluate the reasoning capabilities of Large Language Models (LLMs) in the context of Korean culture. It consists of two subsets: Date and Zodiac, with each subset containing 50 questions, totaling 100 questions across the entire dataset. The dataset is automatically generated via templates and algorithms, requiring models to integrate Korean cultural knowledge during the reasoning process. Its main application focuses on evaluating the reasoning abilities of LLMs, particularly multi-step reasoning tasks related to Korean culture and language. Through this dataset, researchers can better understand model performance in complex reasoning tasks and explore the reasoning capabilities of models in culture-specific contexts.

- 1Multi-Step Reasoning in Korean and the Emergent Mirage延世大学 · 2025年



