Reasoned Safety Alignment (ReSA)
收藏资源简介:
ReSA数据集包含80,000个样例,用于训练大型语言模型(LLMs)以抵御“越狱”攻击。该数据集采用“回答后检查”风格,首先让模型在没有安全约束的情况下尝试直接回答问题,然后基于这个回答进行安全分析。通过这种方式,ReSA旨在提高LLMs对各种“越狱”攻击的鲁棒性,同时保持其在一般任务上的性能,如数学和编码,同时实现低拒绝率。此外,ReSA还配备了安全完成机制,能够在敏感查询(甚至对抗性提示)下提供有帮助和支持的响应,例如自我伤害。这为LLMs在安全领域的发展和应用提供了重要的参考价值。
The ReSA dataset contains 80,000 examples, designed for training large language models (LLMs) to resist "jailbreak" attacks. It adopts an "answer-then-check" paradigm, where models are first permitted to directly attempt answering a query without security constraints, followed by a security analysis based on the generated response. Through this approach, ReSA aims to enhance the robustness of LLMs against various jailbreak attacks, while preserving their performance on general tasks such as mathematics and coding, and achieving a low rejection rate. Furthermore, ReSA is equipped with a safe completion mechanism that can deliver helpful and supportive responses to sensitive queries (even adversarial prompts), such as those related to self-harm. This provides important reference value for the development and application of LLMs in the field of security.




