ItsMaxNorm/SafeDiffusion-R1-dataset

SafeDiffusion-R1提示数据集是一个用于文本到图像生成模型安全训练的提示语料库。该数据集专门为SafeDiffusion-R1模型的GRPO后训练设计，通过闭环式引导奖励来增强Stable Diffusion模型的安全性。数据集包含约235,774条提示，分为10个子集，涵盖良性提示和NSFW（不适合工作场所）负面锚点提示。良性提示用于常规训练，而NSFW提示则专门用于安全研究，作为负面样本来引导模型避免生成不当内容。该数据集旨在支持扩散模型的安全后训练研究，仅限研究用途。The SafeDiffusion-R1 prompt dataset is a training prompt corpora used for safe training of text-to-image generation models. It is specifically designed for GRPO post-training of the SafeDiffusion-R1 model, employing a closed-form steering reward to enhance the safety of Stable Diffusion models. The dataset contains approximately 235,774 prompts across 10 subsets, including benign prompts and NSFW (Not Safe For Work) negative-anchor prompts. Benign prompts are used for general training, while NSFW prompts are intended for safety research as negative samples to steer models away from generating inappropriate content. This dataset is released for safety-research purposes only, supporting the study of safe post-training methods for diffusion models.