CAD-Steps

CAD-Steps is a dataset of intermediate CAD construction states — not just finished models, but every step along the way — for training models on how geometry is actually built.

Most CAD datasets contain only final shapes, so models trained on them learn what objects look like but not how a designer constructs them. CAD-Steps uses OpenCascade to replay construction sequences and capture each intermediate state, producing step-by-step data for training and evaluation.

The dataset is public on Hugging Face, with the generation code on GitHub.