ShiyuanHuang/Text2Space

Text2Space是一个包含20,000个空间推理实例的合成数据集。每个实例将一个2D布局的自然语言描述与同一场景的三个ASCII渲染以及一个关于两个对象相对位置的查询配对。该数据集旨在训练和评估语言和视觉语言模型在空间推理方面的能力。数据集是论文《Learning to Draw ASCII Improves Spatial Reasoning in Language Models》的配套数据集。数据集包含多个字段，如描述、查询关系、标签、ASCII渲染等。生成过程包括采样对象、构建空间关系、渲染ASCII样式等步骤，并在组件数量、关系数量、查询类型等方面进行了平衡。

Synthetic dataset of 20,000 spatial reasoning instances. Each instance pairs a natural-language description of a 2D layout with three ASCII renderings of the same scene and a query about the relative position of two objects. Designed to train and evaluate language and vision-language models on spatial reasoning. Companion dataset for the paper *Learning to Draw ASCII Improves Spatial Reasoning in Language Models*. The dataset includes multiple fields such as description, query relation, label, ASCII renderings, etc. The generation process involves sampling objects, constructing spatial relations, rendering ASCII styles, and is balanced across component counts, relation counts, query types, etc.