Reproduce “CADmium: Fine-Tuning Code Language Models for Text-Driven Sequential CAD Design” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Prashant Govindarajan, Davide Baldelli, Jay Pathak, Quentin Fournier, Sarath Chandar
CADmium enables text-driven sequential Computer-Aided Design by fine-tuning code-specific large language models on a newly curated, high-quality dataset of over 170,000 text-CAD pairs. This method effectively generates precise JSON-formatted CAD sequences, often outperforming specialized baselines in geometric fidelity and structural metrics, with performance scaling positively with model size.
Key insights
- High-quality, multimodal LLM-generated annotations can effectively balance human-like fluency and geometric precision, significantly enhancing the learning for text-to-CAD tasks.
- Fine-tuning general-purpose code-LLMs for text-to-text CAD generation is a viable and scalable strategy, effectively leveraging their pre-trained knowledge to achieve competitive or superior results over specialized architectures.
- Increasing model size consistently improves the robustness and accuracy of generated CAD models, and consistency in annotation style between training and evaluation data is crucial for optimal performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.