CADME-1k: A Multimodal Dataset for CAD Understanding
Name
Spencer Goldstein_Paper.pdf
Size
942.67 KB
Format
Adobe PDF
Checksum (MD5)
e6301dd3869dc27d7af5dffa6080becc
Author(s) •
Goldstein, Spencer
Ghanem, Janelle
Date Issued
October 26, 2026
Abstract
CADME-1k is a multimodal dataset that represents the Computer Aided Design (CAD) creation process visually, geometrically, symbolically, and textually; no existing dataset combines all these components into one. Containing 1,000 engineering assemblies across 12 distinct CAD modalities, CADME-1k enables models trained on this data extract artifacts throughout the CAD process as the dataset is more representative of a complete, manually made CAD model than anything currently available. The dataset allows AI systems to move from recognizing what an object looks like toward understanding how it is constructed and how it can be modified. We will highlight in detail one of the AI-for-CAD tasks CADME-1k supports, constraint extraction from CAD, in this paper.
Subjects
computer-aided design
multimodal dataset
constraint extraction
vision-language models
parametric design
MIT Department
Lincoln Laboratory
Persistent DSpace Link