Video-Reason/VBVR-MultiStep
收藏资源简介:
VBVR-MultiStep是一个用于长视野多步图像到视频(I2V)推理的程序化训练语料库,包含约36万个样本,涵盖36个参数化任务。数据集支持导航、规划、CSP、执行、几何和物理等多种推理家族。每个样本包含初始帧、提示文本、目标帧、真实视频和元数据。数据集格式为Tar.gz分片和Parquet元数据,总大小约164 GB。数据集为完全合成,适用于训练多步推理系统,但不适用于无约束开放世界视频的转移。
VBVR-MultiStep is a ~360k-sample programmatic training corpus for long-horizon multi-step image-to-video (I2V) reasoning, encompassing 36 parameterized tasks. The dataset supports various reasoning families including Navigation, Planning, CSP, Execution, Geometry, and Physics. Each sample includes a first frame, prompt text, final frame, ground truth video, and metadata. The dataset is formatted as Tar.gz shards with Parquet metadata, totaling ~164 GB. It is fully synthetic and intended for training multi-step reasoning systems, though not validated for transfer to unconstrained open-world video.




