Tutu1/LatentConceptMisalignment

该数据集是为研究文本到图像扩散模型中的潜在概念不对齐问题而创建的，包含了不同阶段的实验数据。数据集包括三个JSON文件，分别包含173、90和76个级别5的概念对，这些概念对来源于人类专家总结的4个LC-Mis模式以及GPT-3.5生成的新模式。数据集的级别评分基于2023年10月1日可用的Midjournal图像生成系统，并通过人工评估进行。未来工作将扩展使用的模型类型和版本，并迭代数据挖掘算法以增强和扩大数据集。

This dataset is the official database for studying latent concept misalignment in text-to-image diffusion models. It includes 173, 90, and 76 concept pairs of different levels, derived from initial LC-Mis patterns summarized by human experts and newly generated patterns by GPT-3.5. The ratings are based on a specific version of the Midjournal image generation system and confirmed through human assessments. Future plans include expanding model types and versions, and continuously updating data mining algorithms to enhance the dataset.