Data for: Evol-Preference : Evol-Preference: Automatic Evolution of Preference Data For Safety Alignment

The PKU-Alignment team released the dataset "Beavertails", which focuses on AI safety. We have extended and optimized "Beavertails" to obtain this dataset. Readers can directly use our dataset to train large language models to enhance their usefulness and harmlessness.Training details: 70% for supervised fine-tuning(SFT), 30% direct preference optimization (DPO), training hyperparameters available in Appendix C of the paper.

PKU-Alignment团队发布了聚焦人工智能安全的数据集《Beavertails》。我们对《Beavertails》进行了扩展与优化，得到本次发布的数据集。研究人员可直接使用本数据集对大语言模型（Large Language Model, LLM）进行训练，以提升模型的实用性与无害性。训练细节如下：70%的数据用于监督微调（Supervised Fine-Tuning, SFT），30%的数据用于直接偏好优化（Direct Preference Optimization, DPO）；训练所需的超参数可参见论文附录C。