Intent2CAD

Official dataset for Intent2CAD: How Semantic–Parametric Supervision Shapes Text-to-CAD Generation, accepted at NeurIPS 2026.

Paper · Code

Intent2CAD provides natural language descriptions paired with executable CadQuery programs, along with source–target pairs for intent-preserving CAD editing. Each model has three aligned code representations:

• Primitive: low-level sketch and extrusion operations.
• Semantic: feature-level operations such as rectangles, polygons, slots, fillets, and arrays.
• Semantic–parametric: semantic features with shared parameters and derived expressions.

Datasets

Subset	Splits	Description
generation	train / validation / test	IntentCAD-100K: text descriptions and three aligned code representations.
editing	test	IntentCAD-Edit-1K: editing instructions with source and target programs for each representation.

The editing dataset covers Feature Edit (FE), Consistency Edit (CE), and Derived-Relation Edit (DE), stored as feature, consistency, and dependency.

The editing examples are drawn from the generation test set.

Data Format

Generation

Field	Description
id	Sample identifier
text	Natural language description
code_primitive	Primitive CadQuery program
code_semantic	Semantic CadQuery program
code_semparam	Semantic–parametric CadQuery program

Editing

Each example contains id, category, instruction, and a source–target pair for each representation: source_primitive / target_primitive, source_semantic / target_semantic, and source_semparam / target_semparam.

Usage

from datasets import load_dataset

generation = load_dataset("scottyjb/Intent2CAD", "generation", token=True)
train = generation["train"]
validation = generation["validation"]
test = generation["test"]

editing = load_dataset(
    "scottyjb/Intent2CAD", "editing", split="test", token=True
)

Evaluation

Evaluation annotations are provided in metadata/editing.tsv, matched to examples by id.

They specify the loops checked for edit success, edit types, and relation subtypes.

Models should receive the source program and editing instruction. Target programs and evaluation annotations are used only for scoring.

Citation

@inproceedings{yang2026intent2cad,
  title     = {{Intent2CAD}: How Semantic--Parametric Supervision Shapes Text-to-CAD Generation},
  author    = {Jinbo Yang and Mingyue Yuan and Shikai Jing},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

Acknowledgements

The CAD construction sequences are derived from DeepCAD. Programs use CadQuery.

License

This dataset is released under CC BY-NC 4.0. Source materials retain their original licenses.