A CAD model dataset is useful only when its files contain the information your model needs. A collection of finished solids can support shape analysis. It cannot, by itself, teach a model the order in which a designer made each feature. A dataset of drawings can support drawing understanding, but it needs reliable links to 3D parts if the task is drawing-to-solid reconstruction.

That is why the first question is not “How many CAD files are there?” It is “What is the input, what is the target, and how are they connected?”

Public CAD datasets answer different questions

Source What it provides Best starting point for Limit to check
ABC Parametric curves and surfaces across a large CAD collection Geometric learning and shape reconstruction The project says model copyright remains with the creators. Check its terms for your use.
DeepCAD CAD construction sequences parsed from public Onshape documents Learning or generating modeling steps A sequence dataset is different from a collection of native project files.
Fusion 360 Gallery Human design sequences in a sketch-and-extrude representation Programmatic CAD reconstruction research The dataset license limits use to noncommercial research.
Zero-to-CAD 100K Synthetic CAD programs, rendered views, and STEP outputs Prototyping image-to-CAD and program-generation models Synthetic construction histories do not establish performance on real engineering projects.
OrlaMechCAD-Nano Paired DWG drawings, STEP solids, PDFs, views, and structured relationships for 14 components Inspecting drawing-to-solid and multimodal data pipelines It is a small evaluation sample, released under CC BY-NC 4.0.

These sources are not interchangeable. ABC describes geometry at scale. DeepCAD and Fusion 360 Gallery add design sequences. Zero-to-CAD uses generated programs. OrlaMechCAD-Nano makes the link between a drawing, a solid, and their extracted records explicit.

Look inside one example before choosing a dataset

The file extension tells only part of the story. STEP gives portable solid geometry, but it does not necessarily preserve the original feature timeline or assembly context. A rendered drawing image shows what a reader sees, but structured dimensions and drawing entities can make the same sheet usable for a different task.

For one record, check whether you can answer these questions:

  1. What is the original asset, and what was derived from it?
  2. Can the drawing, model, and metadata be joined with stable identifiers?
  3. Are units, coordinate frames, and transformations documented?
  4. Is the training target actually present, or is the dataset intended only for evaluation?
  5. What does the license permit for this use?

The OrlaMechCAD-Nano card shows one way to expose those links. Each component pairs a DWG drawing and a STEP solid. Its catalogs connect the files to dimensions, face records, measurements, and view metadata. The 14 pairs are useful for checking a parser or defining a task, but they are too few to treat as evidence that a model will generalize.

When a public benchmark is not the whole answer

Public datasets are a good place to compare methods and inspect formats. A production research program may need different coverage: particular part families, native file types, assemblies, revisions, or permission records. A larger file count does not fill those gaps automatically.

Orla works with engineering companies to define collections around a research brief. The aim is to keep native files, linked project context, and source records together where they are available. If your team needs a collection beyond a public benchmark, tell Orla what task, formats, and coverage you need.

For the broader selection problem, read engineering datasets for machine learning.