The key insight: you can get broad geometric coverage for CAD model learning without the compute cost of a million-sample corpus by clustering visual embeddings and picking centroid exemplars. This 100K subset preserves the distributional diversity of the full Zero-to-CAD 1M while cutting storage and iteration time, making it practical for prototyping and benchmarking 3D generation and program-synthesis workflows.
What Sets It Apart
- Embedding-driven curation: samples were rendered from 8 views, encoded with DINOv3, clustered with k-means, and the nearest-to-centroid exemplar from each cluster was selected — so the subset maximizes geometric diversity per sample. This reduces redundant shapes common in random sampling.
- Program-first samples: each entry includes an executable CadQuery construction script (plus STL/STEP exports), enabling research that trains or evaluates models that output parametric CAD programs, not just meshes — so you can study program synthesis, operation vocabulary, and reconstruction fidelity.
- Retrieval-ready: precomputed DINOv3 embeddings and a FAISS IVF-PQ index exist for the full 1M (covering these 100K samples), making similarity search, few-shot retrieval, and retrieval-augmented generation experiments straightforward.
- Resource-aware design: split counts (≈81k train / 9.7k val / 9.8k test) and parquet streaming support are tailored to lower-memory setups and faster iteration compared to the full 1M.
Who it's for and tradeoffs
Great fit if you need a representative, compute-friendly CAD corpus for quick model prototyping, program-synthesis research (CadQuery outputs), retrieval-augmented 3D generation, or benchmarks where geometric diversity matters more than sheer scale. Look elsewhere if you require full coverage of rare part classes, extremely large-scale pretraining, or real-world scanned CAD datasets — the dataset is synthetically agentically synthesized, so domain gap to industry CAD repositories may exist.
Where it fits
This sits between small curated CAD benchmarks (too small for robust generalization) and the Zero-to-CAD 1M (max coverage). Use the 100K for iterations, ablations, and evaluation; move to the 1M when scaling model capacity or training from scratch.
How the subset was selected
Eight-view renders → DINOv3 visual embeddings averaged per part → k-means into 100K clusters → select nearest-to-centroid exemplar per cluster. Each sample retains metadata (operation sequences, timings, face counts) to support analysis of construction complexity and operation vocabulary.
Overall, the dataset is best used when you want program-level CAD supervision and diverse geometry while keeping compute and storage requirements practical. It trades absolute scale for curated coverage and retrieval-ready tooling.
