If you want to sell CAD data for AI models, start with a small inventory, a representative sample, and a clear account of who can license the files. A buyer needs to know what task the archive supports. A folder of drawings, a collection of native parts, and a set of complete assemblies solve different problems.
Match the archive to the buyer's task
| Buyer task | Useful source material | What to check first |
|---|---|---|
| Read or generate technical drawings | Editable DWGs, PDFs, drawing metadata | Are sheets and revisions identified? Can the buyer parse entities and dimensions? |
| Learn part geometry or modeling structure | Native CAD parts, STEP exports, feature history | Are the native files readable? Does each part have the expected history, or only a final solid? |
| Understand an assembly | Assembly geometry, component files, part identifiers, bills of materials | Can a buyer join every component to the assembly and tell repeated instances apart? |
| Link a drawing to a model | Matched drawing and 3D part, with a stable pair ID | Was the link checked, or inferred from similar filenames? |
These are questions to ask, not promises that every buyer will want the same schema. Autodesk's assembly dataset documentation, for example, makes hierarchy, occurrences, geometry, and joints explicit because those fields support assembly research. A drawing-only archive would serve a different task.
Make an inventory before discussing price
For each file family, record the number of files, native format, source project, creation period, approximate size, and whether there are linked drawings or models. Count distinct designs separately from exports and revisions. If one part has a native file, three STEP exports, and two drawings, that is one part with related assets, not six independent designs.
Then inspect a sample from several projects. A useful sample shows the real mix of simple and complex files, broken references, duplicates, and missing metadata. It lets a buyer test whether the proposed training target is actually present. A few handpicked clean files can hide the work needed to use the full archive.
Also record where the files came from. A company may hold project folders that include customer designs, supplier parts, contractor work, and its own original files. Keeping a copy is not, by itself, evidence that the company can grant an AI training license. Our CAD licensing checklist covers that review in more detail.
What a first buyer conversation should contain
You can begin without sending the whole archive. A one-page summary is enough to test fit:
- The engineering domains and projects represented.
- File counts by native format, with duplicates and revisions called out.
- What links exist between drawings, parts, assemblies, and bills of materials.
- The intended licensing scope and any files excluded pending a rights review.
- A representative sample that the buyer can inspect under agreed terms.
This makes a serious conversation possible before anyone negotiates a large delivery. The buyer can state what they need; the owner can see which parts of the archive are worth preparing.
See how distinct CAD collections are described
Orla's DWGAtlas-194K is an existing collection of 194,025 editable DWG drawings, with a 50-drawing sample. Its card states that the source material came from Brazilian engineering companies under exclusive rights. The drawings were prepared from project archives, so the collection is not a promise that every drawing comes with a complete project archive. SLDPRTCorpus-4K is a separate existing collection of 4,878 native SolidWorks parts with saved feature trees. It is a part collection, not an assembly collection.
Those cards show the level of detail a buyer should ask for: file types, source, links, limits, sample, and licensing terms. If you own an engineering archive and want to explore a license, send Orla an inventory and the rights scope you can support. If you are choosing public research data instead, start with our CAD model dataset comparison.