SweEval
收藏资源简介:
SweEval是一个跨语言的基准数据集,用于评估大型语言模型(LLM)在处理敏感语言时的表现。该数据集由多个现实世界场景组成,包括不同的写作风格和语境。数据集包含针对企业和非正式语境的手动创建的指令提示,以及25个来自高资源和低资源语言的咒骂词。这些咒骂词被整合到英语提示中,以评估模型对当地语言细微差别和文化敏感性的理解。SweEval旨在帮助研究人员开发符合道德标准的AI系统,特别是在企业和跨文化环境中。
SweEval is a cross-lingual benchmark dataset designed to evaluate the performance of Large Language Models (LLMs) when handling sensitive language. This dataset comprises multiple real-world scenarios covering diverse writing styles and contexts. It contains manually crafted instruction prompts tailored for corporate and informal contexts, alongside 25 curse words from both high-resource and low-resource languages. These curse words are integrated into English prompts to assess the model's understanding of the subtle nuances and cultural sensitivities of local languages. SweEval aims to assist researchers in developing ethically aligned AI systems, particularly in corporate and cross-cultural environments.

- 1SweEval: Do LLMs Really Swear? A Safety Benchmark for Testing Limits for Enterprise UseOracle AI, Indian Institute of Information Technology Ranchi, TD Securities, Columbia University, Hanyang University · 2025年



