datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
cadevolve
CADEvolve: Creating Realistic CAD via Program Evolution
CADEvolve is a synthetic dataset for CAD reverse engineering built around executable CadQuery programs. It is produced by the CADEvolve pipeline, which evolves a small set of primitives into 8k complex parametric generators, and then expands them into a large corpus of CAD programs (≈1.3M scripts) exercising a broad CadQuery operation set.
Paper: arXiv:2602.16317
Code: GitHub
Repository structure
This… See the full description on the dataset page: https://huggingface.co/datasets/kulibinai/cadevolve.cadevolve
CADEvolve: Creating Realistic CAD via Program Evolution
CADEvolve is a synthetic dataset for CAD reverse engineering built around executable CadQuery programs. It is produced by the CADEvolve pipeline, which evolves a small set of primitives into 8k complex parametric generators, and then expands them into a large corpus of CAD programs (≈1.3M scripts) exercising a broad CadQuery operation set.
Paper: arXiv:2602.16317
Code: GitHub
Repository structure
This… See the full description on the dataset page: https://huggingface.co/datasets/trizhi/cadevolve.cadcad-video-cache
cad-video-cache — pre-encoded latent caches (skip data_process)
Slimmed VAE+T5 latent caches for the CAD video world-model (cad_video_gen). Download +
extract a tar to skip the ~8-9h data_process, then train directly:
train_df.py --dataset_base_path cache_xxx.
What's in a .pth
Each clip is one file cache_*/0/{idx}.pth (~1.3 MB), a 2-tuple of dicts:
[0]["input_latents"] — Wan-VAE video latent (1, 16, T, 60, 60) bf16: 16 channels, T
temporal-latent frames (VAE 4:1… See the full description on the dataset page: https://huggingface.co/datasets/hz6666/cad-video-cache.