XXZ13871076823/Zero-To-CAD-100k

Zero-to-CAD 100K是一个精选的CAD构造序列数据集，包含从Zero-to-CAD 1M全数据集中选取的100,000个几何多样性样本。该数据集专为计算资源有限的研究者设计，通过视觉嵌入（使用DINOv3特征从8个视角渲染模型）、聚类（K-means将嵌入空间分为100K个簇）和选择（每个簇选取最接近中心的样本）过程，确保样本覆盖全数据集中所有零件类型、操作和几何复杂性的分布。数据集分为训练集（81,015个样本）、验证集（9,734个样本）和测试集（9,767个样本）。每个样本包含唯一标识符（UUID）、可执行的CadQuery Python源代码、最终实体的B-Rep面数、CAD操作列表（JSON格式）、操作数量、8个渲染视图（256×256图像）、STL网格文件和STEP文件。数据集支持广泛的CAD操作，包括草图原语（如矩形、圆形、多边形）、3D操作（如拉伸、切割、旋转）、修改（如圆角、倒角）、布尔运算（如并集、交集）和模式（如线性、极坐标）。该数据集适用于快速原型设计、CAD生成模型的评估与基准测试，以及在资源受限环境下微调较小模型。数据集基于Apache License 2.0发布。Zero-to-CAD 100K is a curated subset of 100,000 geometrically diverse CAD construction sequences selected from the full Zero-to-CAD 1M dataset. Designed as an accessible entry point for researchers with limited compute resources, it ensures maximum geometric diversity through a curation process that includes visual embedding (using DINOv3 features from 8 rendered viewpoints), clustering (K-means partitioning into 100K clusters), and selection (choosing the nearest-to-centroid exemplar per cluster). This subset spans the full distribution of part types, operations, and geometric complexity from the 1M dataset. It is split into train (81,015 samples), validation (9,734 samples), and test (9,767 samples). Each sample includes a unique identifier (UUID), executable CadQuery Python source code, number of B-Rep faces in the final solid, CAD operations list (in JSON format), operation count, 8 rendered views (256×256 images), STL mesh file, and STEP file. The dataset covers a broad vocabulary of CAD operations, such as sketch primitives (e.g., rect, circle, polygon), 3D operations (e.g., extrude, cut, revolve), modifications (e.g., fillet, chamfer), Boolean operations (e.g., union, intersect), and patterns (e.g., linear, polar). It is intended for quick prototyping, evaluation and benchmarking of CAD generation models, and fine-tuning smaller models in resource-constrained environments. The dataset is released under the Apache License 2.0.