FairMT-10K
收藏资源简介:
FairMT-10K是由浙江大学创建的多轮对话公平性评估数据集,旨在解决大型语言模型(LLM)在多轮对话中的公平性问题。该数据集包含10280条多轮对话数据,涵盖了刻板印象和毒性两种主要偏见类型,以及性别、种族、宗教等六种偏见属性。数据集的创建过程结合了现有公平性数据集,并通过模板生成多轮对话数据。FairMT-10K主要应用于评估和改进LLM在多轮对话中的公平性,特别是在复杂对话场景中的偏见识别和纠正。
FairMT-10K is a multi-turn dialogue fairness evaluation dataset developed by Zhejiang University, aiming to address the fairness-related issues of Large Language Models (LLMs) in multi-turn conversations. This dataset contains 10,280 multi-turn dialogue samples, covering two primary bias types: stereotypes and toxicity, as well as six bias attributes including gender, race, religion and others. The construction of FairMT-10K integrates existing fairness datasets and generates multi-turn dialogue data via template-based approaches. FairMT-10K is primarily applied to evaluate and improve the fairness of LLMs in multi-turn dialogues, especially for bias identification and correction in complex conversational scenarios.




