Biopharmaceuticals Retrieval-Augmented Generation Evaluation (BRAGE)
收藏资源简介:
BRAGE是一个专为评估大规模语言模型在生物制药领域的查询与参考理解能力(QRUC)而设计的多语言基准数据集。该数据集包含了开放式的生物制药领域查询,旨在解决传统封闭式查询基准存在的问题,如预定义答案无法评估QRUC,以及数据快速演变导致一些查询只能通过LLM的内部知识正确回答。BRAGE由400个生物制药查询和相应的相关与无关参考组成,涵盖了基本生物学、药物开发与设计、临床转化与应用等多个子领域,以评估LLM在利用增补参考回答查询时的性能。
BRAGE is a multilingual benchmark dataset specifically designed for evaluating the Query and Reference Understanding Capability (QRUC) of large language models (LLMs) in the biopharmaceutical domain. This dataset consists of open-ended biopharmaceutical queries, aiming to address the limitations of traditional closed-ended query benchmarks, such as the inability of pre-defined answers to assess QRUC, and the problem that some queries can only be correctly answered using the internal knowledge of LLMs due to the rapid evolution of related data. BRAGE includes 400 biopharmaceutical queries paired with corresponding relevant and irrelevant references, covering multiple subfields including basic biology, drug development and design, clinical translation and applications, to evaluate the performance of LLMs in answering queries by leveraging supplementary references.




