datumo/FinRED
收藏资源简介:
FinRED是首个专门用于评估大语言模型在金融领域安全性的全面红队测试基准。该数据集由12位韩国金融安全研究所的领域专家合作开发,包含5,805个基于真实金融威胁场景的对抗性种子提示。与关注通用危害的通用安全基准不同,FinRED针对领域特定的金融风险,包括监管合规违规、金融欺诈促成和系统性信任侵蚀等需要专业评估的风险。数据集采用新颖的两级金融风险分类法,涵盖5个一级风险类别和26个二级子类别,每个提示都经过专家验证,确保场景真实性和潜在危害的清晰性。
FinRED is the first comprehensive red team testing benchmark specifically designed to evaluate the safety of Large Language Models (LLMs) in the financial domain. Developed collaboratively by 12 domain experts from the Korea Financial Security Institute, the dataset contains 5,805 adversarial seed prompts based on real-world financial threat scenarios. Unlike general-purpose safety benchmarks that focus on universal harms, FinRED targets domain-specific financial risks, including regulatory compliance violations, facilitation of financial fraud, and systemic trust erosion—risks that require professional evaluation. The dataset adopts a novel two-tier financial risk taxonomy, covering 5 primary risk categories and 26 secondary subcategories. Each prompt has been verified by experts to ensure the authenticity of the scenario and the clarity of its potential harms.




