ACIDE/user-vlm-dpo

DPO数据集旨在提高User-VLM模型的公平性和道德对齐性。它包括两个主要的子数据集：BiasVision-DPO和VLM-DPO，每个子数据集都有独特的属性，用于提高模型性能和减少偏见。BiasVision-DPO包含12K条 entries，整合了FairFace和Bias-DPO数据集的数据，特别设计用来解决User-VLM中的性别、种族偏见等不适当问题的偏见。VLM-DPO包含5.4K条通用DPO entries，用于正则化模型、减轻特定偏见的过拟合、防止灾难性遗忘，并增强公平性。

The DPO dataset is designed to enhance the fairness and ethical alignment of User-VLM models. It consists of two primary sub-datasets: BiasVision-DPO and VLM-DPO, each contributing unique attributes to improve model performance and reduce biases. BiasVision-DPO contains 12K entries that integrate data from the FairFace and Bias-DPO datasets, specifically constructed to address biases in User-VLM, particularly concerning sexist, racist, controversial, and inappropriate questions. VLM-DPO comprises 5.4K general-purpose DPO entries, serving to regularize the model, mitigate overfitting to specific biases or dataset artifacts, prevent catastrophic forgetting, and enhance fairness.