Do-Not-Answer

Do-Not-Answer数据集由MBZUAI创建，包含939条不应被负责任语言模型遵循的指令。该数据集通过精细筛选，旨在评估语言模型的安全机制，并推动开源语言模型的安全发展。数据集内容涵盖多种风险类别，包括信息泄露、恶意使用、歧视与毒性语言、错误信息传播及人机交互风险。通过此数据集，研究者可以训练和评估模型在面对潜在危险指令时的响应，从而提高模型的安全性和可靠性。

The Do-Not-Answer Dataset was created by MBZUAI, containing 939 instructions that should not be followed by responsible language models. Through meticulous screening, this dataset is designed to evaluate the safety mechanisms of language models and promote the safe development of open-source language models. The dataset covers multiple risk categories, including information leakage, malicious use, discriminatory and toxic language, misinformation dissemination, and human-computer interaction risks. With this dataset, researchers can train and evaluate models' responses to potentially dangerous instructions, thereby improving the safety and reliability of the models.