3D and CAD datasets for AI. Geometry with a paper trail.
3D and CAD datasets licensed from the engineers, designers, and architects who modeled them, not scraped from a file-sharing site. Meshes, DXF drawings, and point clouds, labeled by people who build things, fingerprinted by shape so a re-exported duplicate cannot hide, and listed on a signed receipt.
What are 3D and CAD datasets for AI?
3D and CAD datasets are collections of geometry, meshes, technical drawings, and scanned point clouds, paired with the labels a model needs to learn from them: what category an object belongs to, which parts it has, what it is for, how big it is, what it is made of, and how a person would describe its shape. They feed text-to-3D generation, CAD generation, part retrieval, simulation, and the spatial reasoning behind robotics and AR.
Geometry is the hardest data type to source honestly. Much of what is downloadable has a licence nobody can interpret, topology nobody cleaned, and semantics nobody wrote. The same part turns up as STL, OBJ, and GLB under three names, and a file hash sees three different models. And the labels that matter, function, tolerances, and manufacturability, are invisible to anyone who has not made the part.
eeko's data practice licenses models and drawings from the designers, engineers, and architects who made them, through our creator licensing program, and commissions new work from the eeko expert network when a category, standard, or level of detail is missing. Every item is fingerprinted by shape, labeled by someone qualified to label it, and delivered with a signed provenance record. It is the provenance discipline we apply to the models we deploy for manufacturing clients.
Mesh, drawing, and point cloud datasets, in the format your pipeline expects.
Geometry as the designer saved it, delivered the way your loader reads it, with the conversion checked for losses.
What we accept and deliver
- Meshes — STL, OBJ, PLY, GLB, and GLTF, with topology checks for non-manifold edges, flipped normals, and holes before acceptance.
- Drawings — DXF technical drawings from engineers and architects, with layers and dimensions intact.
- Point clouds — scanned objects and spaces, captured through our data capture sessions or licensed from scan owners, with personal and location metadata removed.
- Paired items — a drawing matched to its mesh, a render matched to its source, or a text description matched to geometry, for models that learn one from the other.
- Conversion — delivered in the source format or converted to yours, with a diff on vertex count and bounding box so nothing silently collapses.
- Packaging — labels and metadata as JSON Lines, CSV, or Parquet records; WebDataset shards and Croissant metadata for large sets.
Licensed 3D models and CAD data, six ways to fill a spec.
Categories, detail levels, label schema, format, pairing, and review share are written down before the pilot. Then the geometry comes from wherever serves the spec.
CAD datasets for AI labeled by people who build things.
A generalist can tag a mesh as a bracket. Only an engineer can say what it holds, what it is made of, and whether it can be machined.
How geometry gets labeled at eeko
- Schema first — your category tree, part vocabulary, description style, and required fields, agreed and versioned before the pilot.
- Matched expertise — mechanical parts by mechanical engineers, buildings and interiors by architects, consumer objects by industrial designers.
- Source on every label — creator-supplied, AI-suggested and corrected, or expert-written, recorded per field.
- Second expert pass — an agreed share of items reviewed by a second specialist, with disagreement rates reported per batch.
- Pipeline fit — software architects validate the delivered schema against your loader before the first shard ships.
Fingerprinted by shape, not by filename.
The same geometry shows up renamed, rescaled, and re-exported. We check what it is, not what it is called.
3D training data by industry.
From a category list to a sealed dataset.
Fixed scope, fixed price per model, set after the pilot. A real person replies inside one business day.
3D and CAD dataset questions.
What is a 3D and CAD dataset for AI?
A 3D and CAD dataset for AI is a collection of geometry, meshes, technical drawings, and scanned point clouds, paired with the labels a model needs: part segmentation, category and part hierarchies, what each part does, dimensions, materials, and plain-language descriptions of shape and purpose. It feeds text-to-3D, CAD generation, part retrieval, simulation, and spatial reasoning models. A usable one also has clean topology, a licence that permits training, and a record of who made each item.
How much does 3D training data cost?
We price per model against the written specification; larger orders and standing orders get volume pricing. The drivers are whether an item comes from the catalogue or is commissioned, geometric complexity and quality requirements, format conversion, how deep the labels go, which experts have to write them, and exclusivity. A fully labeled CAD assembly from a mechanical engineer costs far more than a tagged decorative mesh. A pilot batch sets the production rate.
Which 3D and CAD file formats do you accept and deliver?
Meshes in STL, OBJ, PLY, GLB, and GLTF, technical drawings in DXF, and point clouds from scanned objects and spaces. Items ship in their source format or converted to whatever your pipeline expects, as the specification states, with labels and metadata as JSON Lines, CSV, or Parquet records. Bigger sets can be bundled as WebDataset shards, with Croissant metadata describing the collection. Conversions are checked for lost geometry before they ship.
How do you detect the same model exported in different formats?
A file hash only catches identical bytes, and the same part saved as STL and again as GLB produces two different files. So every item also receives a shape fingerprint at intake, computed from the geometry itself, which stays the same when a model is renamed, rescaled, re-triangulated, or exported to another format. Both fingerprints are compared against everything already registered, and conflicting ownership claims are held until they are resolved.
Can you commission new CAD models to a specification?
Yes. Mechanical engineers, industrial designers, architects, and CAD technicians in the eeko expert network model to your brief: a category the catalogue lacks, a standard such as a specific thread or fastener family, a level of detail, or paired sets like a drawing with its mesh. Commissioned work is typically licensed exclusively when you need that, and it passes through the same intake, fingerprinting, and second-expert review as catalogue items.
Is downloadable 3D content safe to train on?
Usually not without work. Much online geometry carries unclear or incompatible licences, was re-uploaded by someone other than its author, or was generated procedurally with no record of it. Every item we deliver was licensed for AI training by the person who made it, on stated terms you can read per item, with origin recorded and a signed provenance record. That is what lets you show which models a training set contained.
Explore related data.
Tell us what your model needs to build.
Categories, formats, label depth, and the job the model has to do. A data lead replies inside one business day with a sourcing plan and a pilot proposal. No deck. A ranked list of what to model first.
Experienced within
Markets served.
As an enterprise AI agency, eeko systems delivers production AI systems remote-first across the United States and internationally — including these markets: