GDP-29K
收藏资源简介:
GDP-29K是由中国科学院与阿里巴巴联合构建的大规模几何图表解析数据集,包含20K平面几何(PGDP-20K)和9K立体几何(SGDP-9K)样本,涵盖印刷与手绘两种风格。数据集通过多源真实教育资料收集,经三阶段质量过滤和人工标注,提供几何元素(点/线/面)及语义关系的结构化描述。其统一形式化语言首次实现平面与立体几何的兼容表达,为多模态大模型的几何感知与推理提供关键认知支架,显著提升下游几何问题求解任务的性能。
GDP-29K is a large-scale geometric chart parsing dataset jointly constructed by the Chinese Academy of Sciences and Alibaba. It consists of 20K planar geometry (PGDP-20K) and 9K solid geometry (SGDP-9K) samples, covering two styles: printed and hand-drawn. The dataset is collected from multi-source real educational materials, undergoes three-stage quality filtering and manual annotation, and provides structured descriptions of geometric elements (points/lines/planes) and their semantic relationships. Its unified formal language, for the first time, enables compatible expression of both planar and solid geometry, providing key cognitive scaffolding for geometric perception and reasoning of multimodal large language models, and significantly improving the performance of downstream geometric problem-solving tasks.




