Value Portrait
收藏资源简介:
Value Portrait是一个用于评估大型语言模型(LLM)价值取向的可靠框架。该数据集包含520个独特的查询-响应对,每个都标注了15个心理维度(10个Schwartz价值观和5个五大人格特质)。数据集的建设过程包括从人类-LLM交互数据集中提取查询,使用GPT-4o生成响应,然后由人类参与者根据响应与自身想法的相似程度进行标注。该数据集旨在解决现有基准中存在的价值相关偏差问题,并为理解LLM在真实世界场景中的价值取向提供依据。
Value Portrait is a robust framework for evaluating the value orientations of Large Language Models (LLMs). This dataset includes 520 unique query-response pairs, each annotated across 15 psychological dimensions, namely 10 Schwartz values and 5 Big Five personality traits. The dataset construction workflow involves extracting queries from human-LLM interaction datasets, generating responses using GPT-4o, and then having human participants rate the similarity between the generated responses and their own personal thoughts for annotation. This dataset is designed to mitigate value-related biases present in existing benchmarks, and serves as a reliable basis for understanding the value orientations of LLMs in real-world scenarios.




