CADFS
收藏资源简介:
CADFS是一个大规模参数化CAD模型生成数据集,支持从文本描述和多视图图像生成计算机辅助设计模型。该数据集以FeatureScript程序表示CAD模型,可直接导入Onshape设计环境。数据集旨在训练和评估CADFS模型——一个基于Qwen2-VL-2B微调的多模态大语言模型,专门用于文本到CAD和图像到CAD生成任务。数据内容包含四个主要部分:1)处理后的FeatureScript程序(包括完整精度和缩减精度版本)、自然语言注释(描述几何、拓扑和设计意图)、STEP格式的边界表示几何文件;2)原始源数据,包括未处理的FeatureScript程序、草图元数据、从ABC数据集提取的STEP和STL文件、以及渲染的多视图图像;3)评估基准数据,提供CADFS、DeepCAD和CADParser三个基准测试的最小数据集;4)训练数据,按照两阶段微调策略组织,分为文本输入和图像输入两种模态,并已过滤重复样本和超出上下文长度限制的样本。数据集采用JSONL格式,每个样本包含系统提示、用户输入(文本或图像)和助理输出(FeatureScript代码)。该数据集适用于文本到3D、图像到3D、CAD程序生成和代码生成等研究任务,为基于大语言模型的CAD自动化提供了高质量的训练和评估资源。
CADFS is a large-scale parametric CAD model generation dataset that supports generating computer-aided design (CAD) models from textual descriptions and multi-view images. This dataset represents CAD models via FeatureScript programs, which can be directly imported into the Onshape design environment. The dataset is designed to train and evaluate the CADFS model—a multimodal large language model fine-tuned based on Qwen2-VL-2B, specifically tailored for text-to-CAD and image-to-CAD generation tasks. The dataset content consists of four main parts: 1) Processed FeatureScript programs (including full-precision and reduced-precision versions), natural language annotations (describing geometry, topology and design intent), and boundary representation geometry files in STEP format; 2) Raw source data, including unprocessed FeatureScript programs, sketch metadata, STEP and STL files extracted from the ABC Dataset, as well as rendered multi-view images; 3) Evaluation benchmark data, providing a minimal dataset for three benchmark tests: CADFS, DeepCAD and CADParser; 4) Training data, organized following a two-stage fine-tuning strategy, divided into two modalities: text input and image input, with duplicate samples and those exceeding the context length limit filtered out. The dataset adopts the JSONL format, where each sample contains a system prompt, user input (text or image) and assistant output (FeatureScript code). This dataset is suitable for research tasks such as text-to-3D, image-to-3D, CAD program generation and code generation, providing high-quality training and evaluation resources for CAD automation based on large language models.




