datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
cad-corpus-clean
CAD Corpus — Cleaned & Deduplicated
Unified CadQuery corpus for image→code / text→code CAD model training, built by
cleaning and globally deduplicating 13 source datasets into one corpus.
Rows: 3,755,625 distinct (by exact cq_code)
Quality: 98.6% exec_ok, 98.5% valid_geometry
Schema (17 cols): stem, source_dataset, source_id, source_format, cq_code, nl_description, code_normalized, provenance, license, exec_ok, valid_geometry, bbox_xlen, bbox_ylen, bbox_zlen, num_solids… See the full description on the dataset page: https://huggingface.co/datasets/Hula0401/cad-corpus-clean.cadqa-rl-2000
CAD-QA RL Training Set (2000 samples)
Verifiable-reward RL training rows for CAD geometry reasoning, derived from the
CAD-QA benchmark (release_10k train split, cad_browsecomp difficulty).
Each row is a closed-book question: a CadQuery script + a question about the
resulting geometry. The gold answer is exact-match verifiable.
Files
data/train/ — 2000 rows
data/validation/ — 98 rows (the benchmark's eval-slice rows; excluded
from training — do not train on this… See the full description on the dataset page: https://huggingface.co/datasets/MRiabov/cadqa-rl-2000.