May-apple/VBVR-Reorganized-Image

VBVR-Reorganized-Image数据集是VBVR-Reorganized数据集的图像模式衍生版本，专门用于单图像输入和单图像输出的任务。每个样本由三个文件组成：first_frame.png（作为视觉输入的初始图像）、prompt.txt（经过清理的文本输入提示）和final_frame.png（作为目标输出的最终图像）。数据集旨在支持推理和图像生成任务，涵盖纯推理（Pure_Reasoning）和指令遵循（Instruction_Following）两类任务。总样本数为960,505，包括训练集（960,000个样本）、域内测试集（240个样本）和域外测试集（265个样本）。数据集与原始视频模式版本的主要区别在于：删除了视频文件（ground_truth.mp4）和元数据（metadata.json），仅保留单个提示文件，并排除了10个时间性任务（CLASS_3）。图像模式任务进一步分为两类：CLASS_1（171个任务，直接使用视频模式的最后一帧作为最终图像）和CLASS_2（26个任务，通过每任务渲染器重新生成最终图像）。此外，数据集包含配对变体生成器（如深度翻转和逆向变体）以及极端值翻转变体，用于测试模型对提示的理解而非记忆。数据集适用于图像生成和推理基准测试，使用HuggingFace的datasets库加载，并遵循Apache-2.0许可证。VBVR-Reorganized-Image is an image-mode derivative of the VBVR-Reorganized dataset, designed for single-image-input and single-image-output tasks. Each sample consists of a triple: first_frame.png (visual input), prompt.txt (cleaned text input), and final_frame.png (target output image). The dataset supports reasoning and image-generation tasks, including two main classes: Pure_Reasoning and Instruction_Following. It contains a total of 960,505 samples, split into train (960,000 samples), test_in_domain (240 samples), and test_out_of_domain (265 samples). Key differences from the video-mode parent dataset include the removal of video files (ground_truth.mp4) and metadata (metadata.json), retention of only one prompt per sample, and exclusion of 10 temporal-by-nature tasks (CLASS_3). Image-mode tasks are categorized into CLASS_1 (171 tasks, where final_frame.png is copied verbatim from the video-mode last frame) and CLASS_2 (26 tasks, where final_frame.png is re-rendered from metadata using per-task painters). The dataset also includes paired-variant generators (e.g., depth-flip and inverse variants) and extremum-flip variants to test model understanding of prompts. It is intended for benchmarking image generation and reasoning, can be loaded via HuggingFaces datasets library, and is licensed under Apache-2.0.