Rendra86318/vision-feedback-mix-binarized

该数据集旨在提供大规模的视觉反馈数据，由多个高质量视觉反馈数据集组合而成，包括LLaVA-Human-Preference-10K（9,422个样本）、VLFeedback（80,258个样本）、POVID_preference_data_for_VLLMs（17,184个样本）、RLHF-V-Dataset（5,733个样本）和RLAIF-V-Dataset（83,132个样本）。数据集格式包含样本ID、来源数据集、图像、图像路径和对话内容。该数据集可用于训练视觉奖励模型或进行DPO训练，以对齐视觉大型语言模型。目前数据集未经过滤，但未来会发布过滤版本以提高质量。

This dataset aims to provide large-scale vision feedback data. It is a combination of the following high-quality vision feedback datasets: LLaVA-Human-Preference-10K (9,422 samples), VLFeedback (80,258 samples), POVID_preference_data_for_VLLMs (17,184 samples), RLHF-V-Dataset (5,733 samples), and RLAIF-V-Dataset (83,132 samples). The dataset format includes sample ID, source dataset, image, image path, and conversations. It can be used to train a vision reward model or to conduct DPO training, aligning a vision large language model. Currently, the dataset is unfiltered, but a filtered version will be released soon to improve overall quality.