NaVAB
收藏资源简介:
NaVAB是由香港科技大学等机构提出的一个全面的基准,用于评估大型语言模型(LLM)与五个主要国家(中国、美国、英国、法国、德国)价值观的契合度。该数据集通过一个价值数据提取管道构建,包含话题建模、价值敏感话题筛选和价值评估数据生成三个主要过程。数据来源于八个国家的官方媒体,经过一系列处理,最终生成了用于评估和校准LLM的价值观评估数据。
NaVAB is a comprehensive benchmark proposed by institutions including the Hong Kong University of Science and Technology, which is designed to evaluate the alignment between Large Language Models (LLMs) and the values of five major countries: China, the United States, the United Kingdom, France and Germany. This dataset is constructed via a value data extraction pipeline, which encompasses three core processes: topic modeling, value-sensitive topic screening, and value assessment data generation. The raw data is sourced from official media outlets across eight countries, and after a series of processing steps, it finally generates value assessment data for evaluating and calibrating LLMs.




