WanyueZhang/World2VLM

World2VLM数据集旨在通过世界模型作为教师模型，将空间想象力蒸馏到视觉语言模型（VLMs）中，从而增强其动态空间推理能力。该数据集使VLMs能够在推理时无需外部模拟的情况下，对未来的视图和动作结果进行推理。演示数据集包含基于轨迹的监督样本和8种动态空间推理任务类型的示例。数据集结构包括四个子集，涵盖不同的教师模型和场景类型。每个子集包含一个轨迹束和一个带有结构化监督的tasks_demo.jsonl文件。数据格式和任务套件详细说明了数据构造流程和数据集的关键特征。

The World2VLM dataset is designed to enhance Vision-Language Models (VLMs) with dynamic spatial reasoning capabilities by distilling spatial imagination into VLMs using world models as teachers. This enables VLMs to reason about future views and action consequences without external simulation at inference time. The demo dataset includes trajectory-based supervision samples and examples of 8 dynamic spatial reasoning task types. The dataset structure comprises four subsets covering different teacher models and scene types. Each subset includes a trajectory bundle and a tasks_demo.jsonl file with structured supervision. The data format and task suite are detailed, along with the data construction pipeline and key features of the dataset.