BioHopR
收藏资源简介:
BioHopR是一个专门设计用于评估大型语言模型在生物医学领域进行多跳推理和多答案推理能力的基准数据集。该数据集由PrimeKG知识图谱构建,包括1跳和2跳推理任务,反映了现实世界中生物医学的复杂性。数据集由伦敦大学学院、伦敦国王学院和格拉斯哥大学的Yunsoo Kim、Yusuf Abdulle和Honghan Wu共同创建,旨在填补生物医学领域多跳推理基准的空白。该数据集共包含2,494个唯一的1跳问题和7,633个唯一的2跳问题,总计279,738个答案。数据集可在https://huggingface.co/datasets/knowlabresearch/BioHopR访问。
BioHopR is a benchmark dataset specifically designed to evaluate the multi-hop and multi-answer reasoning capabilities of large language models (LLMs) in the biomedical domain. Constructed from the PrimeKG knowledge graph, this dataset includes 1-hop and 2-hop reasoning tasks that reflect the complexity of real-world biomedical scenarios. It was co-created by Yunsoo Kim, Yusuf Abdulle and Honghan Wu from University College London, King's College London and University of Glasgow, aiming to fill the gap in multi-hop reasoning benchmarks within the biomedical field. The dataset contains 2,494 unique 1-hop questions and 7,633 unique 2-hop questions, with a total of 279,738 answers. The dataset is publicly accessible at https://huggingface.co/datasets/knowlabresearch/BioHopR.

- 1BioHopR: A Benchmark for Multi-Hop, Multi-Answer Reasoning in Biomedical Domain伦敦大学学院、伦敦国王学院、格拉斯哥大学 · 2025年



