ACADEMICarXiv · 2025

Text-to-CadQuery

Trains models (124M–7B parameters) to generate executable CadQuery Python from natural language. Introduces a self-correction loop that revises code when it fails to execute, raising success rate from 53% to 85%. Dataset of ~170K text-CadQuery pairs derived from DeepCAD.

Input
Text
Output
CadQuery Python
Venue
arXiv
Year
2025
Links
Paper →
CAD Arena Status
Scheduled for evaluation on the full 200-prompt benchmark. Results will appear on the leaderboard at launch.