lucio36/APASI-SI-dataset

APASI自我注入（SI）数据集是一种用于减少大型视觉语言模型（LVLMs）中幻觉现象的新型方法。该数据集通过目标LVLM自身向生成的响应中注入幻觉，创建具有不同偏好级别的响应对，进而用于基于DPO的偏好对齐。数据集包括两个子集：SI-23k和SI-130k，分别来源于LLaVA指令调整数据和VisualGenome数据集，用于提供训练LVLMs所需的偏好对。The APASI Self-Injection (SI) Dataset is a novel approach for mitigating hallucinations in Large Vision-Language Models (LVLMs). The dataset uses the target LVLM itself to inject hallucinations into a generated response, creating pairs of responses with varying preference levels for DPO-based preference alignment. It includes two subsets, SI-23k and SI-130k, sourced from LLaVAs instruction tuning data and the VisualGenome dataset, respectively, to provide the preference pairs necessary for training LVLMs.