A CAD assembly dataset brief should say what one record represents, which components belong to it, and which relationships the model needs to learn. “10,000 CAD files” does not answer whether those files are complete assemblies, loose parts, or multiple exports of the same design.
Start with the task. If you want to predict mates, you need ground-truth constraints. If you want to read a bill of materials, you need reliable component IDs and quantities. If you want drawing-to-assembly work, you need drawings linked to the right geometry. The AutoMate project focuses on mating, while Autodesk's Fusion 360 Gallery assembly documentation describes hierarchy, occurrences, geometry, and joints. They show why “assembly data” is not one fixed format.
Define one record and its links
Give the supplier a short schema. A useful starting point is:
| Field | What to ask for |
|---|---|
| Assembly ID | A stable ID for the complete design and every export of it. |
| Component ID | A stable link from each listed part to its model and drawing. |
| Occurrence ID | A separate ID when the same component appears more than once. |
| Geometry | State whether you need assembled STEP, native CAD files, component STEP, or all three. |
| Structure | Ask for parent-child links, transforms, and subassemblies if the task needs them. |
| Relationships | Name the required labels: mates, joints, contacts, fasteners, or just a bill of materials. |
| Drawings | Specify whether drawings exist for the assembly, each component, or both. |
| Provenance | Record source project, revisions, permitted uses, and any excluded files. |
Do not assume that a bill of materials contains a complete assembly graph. A BOM can list parts while leaving position, orientation, and motion constraints unstated. Likewise, an assembled STEP model may show final geometry without native design history.
Add acceptance checks before discussing scale
Ask for ten records from different projects and inspect them before agreeing to a full delivery. For each record, try to open the assembled model, resolve every component ID, check units and coordinate frames, compare the drawing with the geometry, and count missing or duplicate files. If the brief requires quantities or mates, verify those fields are populated in the sample. Run the same checks automatically on the full delivery.
Specify a split by source project or design family when evaluating a model. Random file splits can put revisions or related components on both sides of the test, making results hard to interpret. Ask how duplicates, near duplicates, and revisions will be labeled.
A short request you can send
We are building a model for [task]. We need [number] distinct mechanical assemblies in [domain], with one stable assembly ID and linked component IDs per record. Please include [formats], [required relationships], and [drawings or BOM fields]. We need documentation for units, revisions, source projects, rights, and missing values. Start with ten representative records so we can test joins and coverage.
Replace the brackets with requirements the model actually needs. That gives a data provider something it can price and validate.
Orla's AssemblyGraph-10K is an example of a linked assembly offering. Its published sample connects assembled STEP geometry, component models, drawings, renders, and a BOM through component IDs. The full collection is sourced on request, and 10,000 assemblies is an estimate, not ready inventory. The card also says the BOM quantity field was blank in the records reviewed so far. It does not claim mate labels. Review the sample and send Orla your task and required fields if that structure is close to your need.
For a broader look at public sources, read our CAD model dataset comparison.