Datasetpublic

Ayt1da/tube-cad-synthetic

Synthetic Tube CAD 22,800 generated CAD parts. Four classes. Original geometry and labels. Identify round, square and rectangular tubes, plus confusing non-tubes. The parts include bends, holes, slots and cut ends. Browse the files · W&B experiments · The story behind the dataset Choose a collection Main collection Targeted collection Start here if… You want the broad original dataset You want additional difficult variations Parts 20,000 2,800… See the full description on the dataset page: https://huggingface.co/datasets/Ayt1da/tube-cad-synthetic.

sourceHugging Facecc0-1.0updated 6d agoView on Hugging Face

1likes125downloads

Dataset Card

Synthetic Tube CAD

22,800 generated CAD parts. Four classes. Original geometry and labels.

Identify round, square and rectangular tubes, plus confusing non-tubes. The parts include bends, holes, slots and cut ends.

Browse the files · W&B experiments · The story behind the dataset

[image]

Choose a collection

	Main collection	Targeted collection
Start here if…	You want the broad original dataset	You want additional difficult variations
Parts	20,000	2,800
Train / validation / test	14,000 / 3,000 / 3,000	1,920 / 480 / 400
Cached points per part	2,048	4,096
Download	Open main/	Open targeted/

Each collection is balanced across the four classes. Their point caches use different sampling methods; use the STEP files if you want to create one consistent representation. Read the technical details.

What to download

—Want the actual parts? Download the .tar.gz archives in your chosen collection. They contain both STEP geometry and NPZ point clouds with surface normals.
—Want labels or shape details? Start with train.csv, val.csv or test.csv. Each row identifies its archive and files.
—Want generation settings and seeds? Use that collection's manifest.json.

The numbered archives are smaller download chunks, not different dataset versions. Download all the chunks for the split you need, then extract them into the same directory. The combined archive download is about 1.30 GB.

text

main/       Original 20,000 parts, metadata and manifest
targeted/   Additional 2,800 parts, metadata and manifest
previews/   Example images
docs/       Technical details and release summary

Use it from Python

Read the metadata without downloading every CAD file:

python

from datasets import load_dataset

data = load_dataset("Ayt1da/tube-cad-synthetic", "main")
print(data["train"][0])

Use "targeted" for the other collection. To download everything, including geometry:

python

from huggingface_hub import snapshot_download

snapshot_download(
    "Ayt1da/tube-cad-synthetic",
    repo_type="dataset",
    local_dir="tube-cad-synthetic",
)

Extraction, point-cloud format and checksum instructions →

Labels

ID	Label	Example
0	`not_tube`	Solid rod, channel, plate, blind block or housing
1	`round_tube`	Hollow round section
2	`square_tube`	Hollow square section
3	`rectangular_tube`	Hollow rectangular section

The label describes the underlying section, including supported bends and cuts—not the bounding box of the finished part.

Connected to W&B

The original dataset versions and experiments remain available:

—[Main dataset, v0](https://wandb.ai/ayt1da-org/tube-classifier/artifacts/dataset/tube-data-b817913e72b6/v0) · [Targeted dataset, v0](https://wandb.ai/ayt1da-org/tube-classifier/artifacts/dataset/tube-targeted-v4/v0)
—4.94M geometry-model training
—26-view visual comparison · Qwen3.8-27B comparison

Scope and license

Synthetic research data, not a real-world manufacturing benchmark. Train and test shapes share generator families, and these test sets have already been inspected. Strong scores here do not establish accuracy on unfamiliar CAD.

No company data or imported FreeCAD bodies are included. The separate 160-part visual test is also outside this release. See coverage and limitations before comparing results.

Released by Ayt1da under [CC0 1.0](LICENSE). Attribution is not required. Original STEP files, point-cloud contents and split assignments are unchanged by the folder reorganization.