introvoyz041/Zero-To-CAD-1m

Zero-to-CAD 1M是一个大规模数据集，包含100万个可执行的、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环境中操作生成，覆盖了广泛的CAD操作词汇，包括布尔运算、圆角、倒角、抽壳、放样、扫描和模式等，超越了传统的草图拉伸工作流。每个样本都是有效的CadQuery Python代码，可运行并生成几何体，具有命名参数、逻辑构建顺序和人类可读的代码，并经过多阶段几何验证（拓扑、连续性、复杂性、导出）。生成过程采用代理搜索框架，LLM在循环中与CadQuery执行环境交互，使用工具执行验证、查找文档和精确语法查找，迭代生成、执行和修复代码，直到验证成功。数据集包含999,633个序列，分割为训练集（979,633个样本）、验证集（10,000个样本）和测试集（10,000个样本）。数据字段包括唯一标识符、CadQuery源代码、面数、操作列表、渲染视图图像、STL和STEP文件等。适用于训练CAD序列模型、图像到CAD重建、CAD程序理解和基准测试等用途。

Zero-to-CAD 1M is a large-scale dataset of 1,000,000 executable CAD construction sequences generated by an LLM operating inside a feedback-driven CAD environment. It covers a broad vocabulary of CAD operations including booleans, fillets, chamfers, shells, lofts, sweeps, and patterns, going beyond prior sketch-and-extrude workflows. Each sample is executable CadQuery Python code that runs and produces geometry, with named parameters, logical construction order, human-readable code, and multi-stage geometric validation (topology, contiguity, complexity, export). The generation process uses an agentic search framework where an LLM iteratively generates, executes, and repairs code in a loop with a CadQuery execution environment, equipped with tools for validation and documentation lookup. The dataset contains 999,633 sequences split into train (979,633 samples), validation (10,000), and test (10,000). Data fields include unique identifier, CadQuery source code, face count, operation list, rendered view images, STL, and STEP files. It is intended for training CAD sequence models, image-to-CAD reconstruction, CAD program understanding, and benchmarking.