A CAD dataset's licence is not the licence of the models inside it
· 6 min read · Written with an AI assistant
The ABC dataset is released under the MIT License, yet its own archive record says the copyright of the CAD models belongs to their creators. It points to a clause in Onshape's terms for the details. Here is how to read that, and the proven
Most public CAD datasets ship with one licence file, and most teams read it once. That is the mistake. The licence on a dataset and the rights in each model inside it can be two different things, and the best-known public CAD dataset says so on its own archive record.
That dataset is ABC. Its record describes it as a collection of one million CAD models for research into geometric deep learning methods and applications. If your team is building training data for CAD AI, or being sold a tool trained on it, ABC is the obvious place to learn to read the small print.
Two licences in one download
The NYU Faculty Digital Archive record for ABC states that the dataset is released under the MIT License. Many people stop reading there.
The same record also states that “The copyright of these CAD models is owned by their creators”. For licensing details, it points to section 1.g.ii of Onshape’s Terms of Use. It also credits Onshape with providing the CAD models.
Read those statements together. The plain reading is this:
- The MIT licence covers what the researchers released.
- Each model stays the property of whoever designed it.
- Rights in the models are governed by the terms of the platform where they were shared.
This is not legal advice, and a lawyer should review any commercial use. But you don’t need to be a lawyer to see that a record naming two sets of terms is asking you to read both.
The failure mode is easy to picture. Someone searches the download for a LICENSE file, finds MIT and ticks the box on a supplier questionnaire. The second set of terms sits on another website, behind a clause number.
We have not quoted section 1.g.ii here because we have not checked its current wording for this piece. That gap is the point. The answer to “can we train a commercial model on this?” lives in a document the dataset does not contain. Platform terms also change over time. The version that applied when a model was shared may not be the version online today.
Why provenance weighs more for CAD than for photos
A photograph of a bracket shows what the bracket looks like. A parametric model of a bracket defines it: its dimensions, its feature order, and the relationships the designer chose to constrain. Those choices are design decisions, and a model trained on them is learning design decisions, not just shapes.
Three things follow.
- Similarity is easier to judge. Suppose a generated part closely mirrors a training model, with the same hole pattern, the same wall thickness and the same fillet strategy. The question of whose design it is becomes concrete in a way it rarely is for an image.
- Metadata leaks. CAD files often carry part numbers, configuration names, custom properties and material callouts. These can identify a company or a supplier even when the geometry looks generic.
- Engineers are accountable for geometry. If a design is challenged, an engineering lead needs to say where it came from. “The tool suggested it” is not traceability.
ABC’s models came from a public CAD platform, and researchers then assembled and released them. That chain is normal for public CAD data. It is also why the record for the collection and the terms for its contents can differ.
What makes ABC worth the effort
None of this is an argument against ABC. It is useful data, which is exactly why it gets reused.
Each model is described as a collection of explicitly parametrised curves and surfaces. The authors say this gives ground truth for differential quantities, patch segmentation, geometric feature detection and shape reconstruction.
To see why that matters, compare a mesh exported to STL. The mesh has lost the fact that a face was a cylinder of a given radius. It keeps only triangles that approximate one. A parametric description keeps the exact surface type, so normals, curvature and the boundary between faces are known exactly. Those are what you need to tell a fillet from a chamfer, or a drilled hole from a boss.
The authors also say that sampling the parametric descriptions lets users generate data in different formats and resolutions. You can produce point clouds, meshes or other representations at the density your method needs. The paper’s example use case is a large benchmark of surface-normal estimation methods.
The dataset is split into chunks of 10,000 objects. The record cited here is for Chunk 0080, which is linked to CVPR 2019 and was deposited in the archive on 20 April 2019.
Worked example: the provenance record before training
A design-automation lead wants to train a classifier that tags holes, fillets and chamfers in incoming supplier models. ABC’s feature ground truth makes it a natural starting point, and one chunk is a manageable first experiment.
Before anything enters the training pipeline, the lead writes a provenance record. It is kept in the same repository as the training code:
dataset: ABC, Chunk 0080 (10,000 objects)
source record: NYU Faculty Digital Archive, handle 2451/44408
deposited: 2019-04-20
dataset licence: MIT
content rights: copyright owned by model creators
governing terms: Onshape Terms of Use, s.1.g.ii (wording not yet reviewed)
attribution: CVPR 2019 paper; Onshape credited as model provider
intended use: internal R&D or commercial product? <- decide
status: BLOCKED until governing terms reviewed
Every line above the last two comes straight from the archive record. The last two are decisions, and the BLOCKED status is deliberate. A training run is easy to start and hard to undo, so the gate belongs before the run, not after.
The questions behind the record work for any dataset, not just ABC:
- What does the dataset licence attach to? Is it the collection, the processing code, or the contents?
- Who owns the contents, and does the record say so?
- Which third-party terms govern the contents, and which version applied when the data was collected?
- What attribution is required, and where will it appear?
- Is the use research or commercial? Does the answer change if the trained model ships inside a product?
- Can you remove an item if its owner asks? Keep chunk and object identifiers through every preprocessing step so that you can.
The last question will feel familiar to anyone who has traced a part through an E-BOM. If you cannot say which source objects went into a training set, you cannot answer a removal request. Nor can you answer an auditor.
What this means for CAD and engineering tooling
For teams buying CAD AI, “trained on open data” is not an answer. Ask which datasets were used. Ask which terms govern the models inside them. Ask whether the vendor can trace a training sample back to its source. A vendor who can only name the dataset licence has answered half the question.
The same logic runs the other way, towards your own designs. Any model you put on a hosted platform is governed by that platform’s terms. Read the clause on public or shared documents before anything leaves draft. A bracket made public today could end up in a crawl you never agreed to.
This is part of why we built AI-CAD, the engineering copilot for SOLIDWORKS that Bluecoderhub is developing, around self-hosting. It runs on our own fine-tuned, self-hosted model with no external AI services, so customers’ CAD files and engineering data stay inside their own environment. AI-CAD is still a prototype. We are also building datasets for training CAD AI, so every question in the checklist above applies to our own work too.
Read it twice
The ABC record is unusually clear. It names the dataset licence, names the model owners and points to the clause that governs the rest. The risk is not hidden terms. The risk is a team that reads one line and stops. Read the licence once for the dataset and once for what is inside it, and write down what you found before the first epoch runs. If you are working through the same questions on your own training data, follow the journal or get in touch.