New Multimodal CAD Dataset Released for AI Design Research

Researchers have introduced FllumaOne, a novel code-native multimodal CAD dataset designed to advance editable computer-aided design (CAD) research. Unlike traditional datasets that might only provide final geometry, FllumaOne emphasizes the ordered construction history and modeling operations, which are crucial for understanding how a part can be edited. The dataset's models are generated by executable Python programs within the Flluma CAD system, ensuring a high degree of fidelity and editability. The primary release, FllumaOne-100K, comprises 100,000 validated samples spanning four levels of template complexity. Each sample in the dataset meticulously aligns its Python program with a structured feature tree, an intermediate representation optimized for training, as well as STEP geometry, a surface point cloud, natural-language descriptions, metadata, and eight canonical visible-edge renderings. This rich multimodal structure provides a comprehensive view of each CAD model. Rigorous validation processes were applied, including checks for kernel geometry, solid validity, and export success, with detailed reports on modality completeness and duplicate tests. A baseline model, Qwen2.5-Coder-1.5B LoRA, trained on 80,000 samples, achieved impressive results on the held-out test set, demonstrating high Python syntax validity, Flluma build success, and STEP-export validity. This dataset is expected to support a wide range of research areas, including conditioned CAD reconstruction, executable program synthesis, feature-tree prediction, and editable reverse engineering.