mrmikeymarks/Zero-To-CAD-1m

Zero-to-CAD 1M 是一个大规模数据集，包含100万个可执行且可解释的CAD构建序列，完全无需真实世界数据合成生成。该数据集由大型语言模型在反馈驱动的CAD环境中通过代理搜索框架生成，每个样本都是有效的CadQuery Python代码，能够运行并生成几何体，且经过多阶段几何验证（包括拓扑、连续性、复杂性和导出检查）。数据集覆盖了广泛的CAD操作词汇，包括草图基元（如矩形、圆形、多边形）、3D操作（如拉伸、切割、旋转、放样、扫描）、修改操作（如圆角、倒角、壳、偏移）、布尔运算（如并集、差集、交集）、模式（如线性模式、极坐标模式、镜像）以及特征（如孔、螺纹、肋）。数据集中包含来自65个类别的零件，涵盖结构组件、旋转元件、外壳和紧固硬件等。数据集旨在用于训练CAD序列模型、图像到CAD重建、CAD程序理解和基准测试。

Zero-to-CAD 1M is a large-scale dataset of 1,000,000 executable and interpretable CAD construction sequences synthesized entirely without real-world data. It is generated by an LLM operating inside a feedback-driven CAD environment using an agentic search framework. Each sample consists of valid CadQuery Python code that runs and produces geometry, with multi-stage geometric validation (including topology, contiguity, complexity, and export checks). The dataset covers a broad vocabulary of CAD operations, including sketch primitives (e.g., rect, circle, polygon), 3D operations (e.g., extrude, cut, revolve, loft, sweep), modifications (e.g., fillet, chamfer, shell, offset), booleans (e.g., union, cut, intersect), patterns (e.g., linear patterns, polar patterns, mirror), and features (e.g., holes, threads, ribs). It includes parts from 65 categories spanning structural components, rotational elements, enclosures, and fastening hardware. The dataset is intended for training CAD sequence models, image-to-CAD reconstruction, CAD program understanding, and benchmarking.