Op-CAD Dataset

Official dataset release for the ICML 2026 paper Op-CAD: Benchmarking and Investigating Operation-oriented CAD Generation.

Op-CAD is a large-scale multimodal dataset for operation-oriented CAD generation. It contains 128,449 modeling operations across four operation types—extrude, revolve, chamfer, and fillet—and five data modalities. The benchmark represents both individual modeling steps and their position in a complete procedural history.

Resources:

• Paper: https://proceedings.mlr.press/v306/bai26a.html
• Project page: https://baiyixue01.github.io/op-cad/
• Code: https://github.com/Baiyixue01/op-cad
• Data rendering code: https://github.com/Baiyixue01/data_render
• Model: https://huggingface.co/Biabai/op-llama-8b

Contents

• shard_work_shards_op_image/: operation-oriented rendered image shard metadata.
• step_files_pc_2048_normalized/: normalized 2048-point point clouds for STEP files.
• op_oriented_step_pc_2048_normalized/: operation-oriented 2048-point point clouds.
• op-cad-data/: OP-CAD CSV/JSON metadata from the code repository.

Download

Download the complete release with the Hugging Face CLI:

huggingface-cli download Biabai/op-cad \
  --repo-type dataset \
  --local-dir data/op-cad

Large assets use Hugging Face Xet storage. The GitHub repositories contain the scripts used to generate, render, and evaluate the data.

Tasks and evaluation

The paper studies operation-oriented CAD code generation by separating geometric modeling from spatial localization. It also introduces Chamfer/Fillet Intersection over Union (CF-IoU) for evaluating edge selection in chamfer and fillet operations. Reproducible evaluation scripts and prompts are available in the code repository.

Citation

@InProceedings{pmlr-v306-bai26a,
  title     = {Op-{CAD}: Benchmarking and Investigating Operation-oriented {CAD} Generation},
  author    = {Bai, Yixue and Gu, Yufei and Xie, Zeke},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  pages     = {5245--5264},
  year      = {2026},
  volume    = {306},
  series    = {Proceedings of Machine Learning Research},
  publisher = {PMLR},
  url       = {https://proceedings.mlr.press/v306/bai26a.html}
}