RoomSpace

RoomSpace数据集由利兹大学计算学院和阿兰图灵研究所创建，旨在评估语言模型在空间推理方面的能力。该数据集包含10,000个房间的3D模拟数据，每个房间都有多种对象及其空间关系。数据集通过详细的自然语言描述，模拟真实世界的复杂性，包括拓扑、方向和距离关系。创建过程中，对象的布局和空间关系被精心设计，以生成故事、问题和答案。RoomSpace数据集主要用于解决语言模型在空间推理中的多跳推理和不同视图描述解释的挑战，为未来的改进提供了方向。

The RoomSpace Dataset was developed by the School of Computing at the University of Leeds and the Alan Turing Institute, with the goal of evaluating the spatial reasoning capabilities of language models. It encompasses 3D simulation data for 10,000 rooms, each featuring a variety of objects and their corresponding spatial relationships. To simulate the complexity of real-world scenarios, the dataset utilizes detailed natural language descriptions covering topological, directional, and distance-based spatial relationships. During its development, object layouts and their spatial relationships were meticulously designed to generate stories, questions, and corresponding answers. The RoomSpace Dataset is primarily intended to address the challenges of multi-hop reasoning and the interpretation of descriptions from diverse perspectives in language model spatial reasoning, providing guidance for future model improvements.