CAD and PCB engineering files with revision history for AI training

A CAD and PCB engineering dataset is a set of real product designs in the formats engineers work in: native 3D models and assemblies, 2D drawings, schematics and board layouts, with their revision history, engineering change orders and bills of materials. SourceX sources these files where a buyer requests them, from manufacturers and engineering firms that own the designs, and screens each project for customer ownership and export controls such as ITAR and EAR before licensing.

Request this dataset Email the data team

Dataset manifest

Sourced to your spec

What it is: Native CAD and ECAD design files with revisions, change orders and BOMs
Typical systems: SolidWorks, Creo, NX, Fusion, Altium, KiCad, Cadence, with PDM or PLM vaults
Typical history: Varies by partner and project; full revision history only where vaults kept it
Modality: Native and neutral CAD files, PDF drawings, structured BOM and ECO records
Delivery formats: Agreed per order; native files with STEP, IPC-2581 or Gerber exports
Preparation: Title blocks, file properties and PDM user names de-identified; library parts flagged
Licensing: Partner-owned or customer-authorized designs only; field-of-use limits can apply
Availability: Sourced where a buyer requests them, after export-control screening; not guaranteed

Request this dataset

What a delivery contains

Fields vary by source system and are fixed per order. A typical delivery includes:

Field	Type	What it holds
item_id	string	Pseudonymous ID for a part, assembly, drawing or board, stable across files, BOMs and change orders.
item_type	enum	Part, assembly, drawing, schematic, PCB layout, PCB assembly or cable assembly.
authoring_tool	object	The CAD or ECAD application and version that saved the native files.
files	array	Each file's role and format, marked native or neutral — STEP, Parasolid, Gerber, ODB++, IPC-2581 or PDF.
structure	object	Assembly tree with configurations and mates, or for boards the schematic sheets, layer count and stackup.
properties	object	Material, finish, mass, tolerancing standard and other custom properties, with names and company details removed.
revisions	array	Released revisions with lifecycle state, dates and the ECO behind each, plus working versions where the vault kept them.
ecos	array	Change requests and orders — reason, affected items, before and after revision, stock disposition, effectivity and approver roles.
bom	array	Find numbers or reference designators, part IDs, quantities, make or buy, manufacturer part numbers, alternates and do-not-populate flags.
constraints	object	Board design rules such as net classes, clearances and impedance targets, or a part's critical drawing tolerances.
manufacturing_outputs	array	Drill and pick-and-place files, assembly drawings and CAM setups, each matched to its source revision.
third_party_content	array	Vendor part models and library footprints or symbols, flagged by source so their terms can be checked.
screening	object	Ownership basis and export-control screening result for the project, based on the partner's classification records.

Example record

{
  "item_id": "pcba_6c41",
  "item_type": "pcb_assembly",
  "description": "Motor controller board, 4-layer",
  "authoring_tool": { "name": "Altium Designer", "version": "23" },
  "files": [
    { "file_id": "f_a12", "role": "schematic", "kind": "native", "ext": ".SchDoc", "sheets": 6 },
    { "file_id": "f_a13", "role": "layout", "kind": "native", "ext": ".PcbDoc" },
    { "file_id": "f_a14", "role": "fabrication", "kind": "neutral", "format": "IPC-2581" },
    { "file_id": "f_a15", "role": "fabrication", "kind": "neutral", "format": "Gerber X2 + Excellon" },
    { "file_id": "f_a16", "role": "board_3d", "kind": "neutral", "format": "STEP AP214" }
  ],
  "structure": { "layers": 4, "stackup": "stk_4l_1.6mm", "parent": "asm_drive_2e0" },
  "lifecycle": { "state": "released", "revision": "C" },
  "revisions": [
    { "rev": "A", "released": "2022-03-02", "eco": null, "summary": "Prototype build" },
    { "rev": "B", "released": "2022-08-19", "eco": "eco_0231" },
    { "rev": "C", "released": "2023-04-07", "eco": "eco_0318" }
  ],
  "ecos": [
    {
      "eco_id": "eco_0318",
      "requested_by_role": "test_engineer",
      "reason": "Gate driver brown-out at motor start when the supply sags",
      "changes": ["C41 10uF -> 47uF", "U7 swapped for lower-dropout LDO", "TP14 added on VDRV"],
      "affected_items": ["pcba_6c41", "pcb_6c40", "sch_6c3f"],
      "disposition": { "wip": "rework", "finished_stock": "rework", "fielded_units": "no_action" },
      "effectivity": "next_build_lot",
      "approver_roles": ["hardware_lead", "manufacturing_engineer", "quality"]
    }
  ],
  "bom": [
    { "refdes": ["U7"], "part_id": "prt_ldo_0091", "mfr": "[MANUFACTURER]", "mpn": "[MPN]", "qty": 1 },
    { "refdes": ["C41"], "part_id": "prt_cap_4471", "value": "47uF 25V X5R 1210", "qty": 1, "alternates": 2 },
    { "refdes": ["R12", "R13"], "part_id": "prt_res_0221", "value": "10k 1% 0402", "qty": 2 },
    { "refdes": ["J3"], "part_id": "prt_con_0310", "qty": 1, "dnp": true }
  ],
  "third_party_content": [ { "item": "J3 footprint and 3D model", "source": "vendor_library" } ],
  "screening": { "ownership": "partner_owned", "itar": false, "ear": "EAR99 per partner records" }
}

Synthetic record for illustration. Field names, structure and format are agreed per order.

What AI teams use it for

Train generative CAD on design intent

Native files keep sketches, constraints and feature order, so models learn how engineers build a part, not just its final shape, and can produce output an engineer can still edit.

Build PCB design assistants

Schematics paired with their layouts, design rules and stackups show how a circuit becomes a routed board, including placement and routing choices made for noise, heat and assembly.

Model engineering change

ECOs linked to before-and-after revisions show what changed, why, and which parts, drawings and BOMs it touched, the signal change-impact and design-review agents need.

Extract and check BOMs

BOMs tied to their assemblies and schematics give ground truth for BOM extraction from drawings, part substitution and design-to-BOM consistency checks.

Evaluate engineering agents on unpublished designs

Proprietary designs that have never been posted online make contamination-free test items for drawing interpretation, CAD question answering and design review.

Use-case guides: Robotics and embodied AI, Private evaluation sets

What makes this data valuable

Rebuildable native files

Models that open and rebuild without errors keep the feature tree, constraints and design intent.

Revisions tied to change orders

An ECO with a stated reason behind each revision turns a geometry or netlist diff into a labeled change.

Consistent product structure

Assemblies, drawings, BOMs and fabrication outputs that agree let you cross-check every view of a design.

Designs that reached production

Released, built designs carry tolerances, manufacturing notes and post-release fixes that unbuilt concepts never get.

Working versions between releases

Check-ins between releases show how a design converged, including approaches dropped before release.

Managed libraries and naming

Consistent part numbering, layer names and component libraries make a large archive much easier to parse.

The released file shows the result; the history shows the engineering

An engineering archive is most valuable for its record of how designs changed and why, which no final released file shows. Public CAD collections come largely from openly shared models: many are single parts, few carry revision history, and almost none record why a change was made or whether the design was ever built.

A manufacturer's vault holds the whole arc: concepts and working versions, a design review, a release, then change orders triggered by test failures, supplier substitutions, obsolete components, field returns or cost reduction. Each change order pairs a problem statement with a before-and-after design and a list of everything it touched — a supervised signal for models that propose or review changes. BOMs and drawings tie the geometry to manufacturing reality: the material actually specified, the tolerance that mattered, the part swapped in when the original went end-of-life.

Scoping around the representation you train on

A CAD dataset is scoped around the representation your model will consume, because that decides which files, tools and checks matter. Models that learn feature sequences need native files, which usually take the authoring application, its API or a licensed translator to read, since each CAD system stores its history in its own proprietary format. Models that work on boundary representations can use STEP or Parasolid exports. Board-level models need the schematic, netlist and design rules alongside the layout, and KiCad's plain-text files are easier to parse than most.

Engineering data also comes with checks that need no human grader. A model rebuilds or it fails, an assembly has interferences or it does not, a board passes its design rule and electrical rule checks or it does not, and a BOM matches its schematic or it does not. These give verifiable signals for evaluating generated designs and edits, while the change orders add human judgment on top.

Naming the target representation, tools and versions, industries and depth of history in the request lets candidate projects be screened and tested against what you will actually do with them.

What to check before licensing

Confirm each project's ITAR or EAR classification and who made it. Releasing controlled technical data to a foreign person can count as an export even inside the US, so plan who gets access.
Ask how ownership was established per project, and check samples for customer names, part numbers or title blocks suggesting a design was made for someone else.
Check how much of each assembly or board is third-party content, such as vendor part models and library footprints, and what the terms of those libraries allow.
Open sample native files in the tool versions you run, since newer files often will not open in older releases, and check that models rebuild and references resolve.
Confirm that neutral exports, drawings and BOMs were generated from the same revision as the native files they accompany.
Review title blocks, custom properties, PDM check-in comments and PCB silkscreen for names, logos, customer part numbers and addresses.
Ask whether any files were created under educational or personal-use CAD licenses, whose terms can restrict commercial use.
Agree field-of-use terms early. A partner licensing its product designs may want limits on generating close copies of its products.

How licensing works through SourceX

1
Define
Send the domain, modality, volume, format, timeline and permitted use you need.
2
Source
SourceX identifies businesses that hold matching data and are open to licensing it.
3
Qualify
Fit, rights and quality are checked, and you review samples before committing.
4
License
Scope, permitted use, exclusivity, price and obligations are agreed in writing.
5
Deliver
Approved data is prepared, de-identified where required and transferred securely.

Questions buyers ask

Can I license proprietary CAD and PCB design files for AI training?

Yes, where the company that owns the designs agrees. SourceX sources these files where a buyer requests them: you describe the domain, file types, tools and history you need, and SourceX approaches manufacturers and engineering firms with matching designs. Each project is checked for ownership and export-control status before it is offered. Supply depends on which firms take part, so it is not guaranteed.

Will I get native files or only STEP and Gerber exports?

Native files where the partner can provide them, usually with neutral exports alongside. Native SolidWorks, Creo, NX or Fusion files keep the feature tree, sketches and constraints; a STEP export keeps solid geometry and assembly structure but usually not how the model was built. Gerber files describe each board layer's image and may carry little or no netlist or component data, which the native design, ODB++ or IPC-2581 hold.

How do export controls apply to CAD and PCB data?

Engineering design files can be controlled technical data. In the US, technical data for defense articles falls under the ITAR, and technology for many commercial and dual-use items falls under the EAR, where classification, destination, end user and end use decide whether a license is needed. The EU and other jurisdictions have their own dual-use controls. Projects are screened against the partner's classification records, designs that would need a license are left out, and your own export compliance team should review the scope too.

Is the full revision and change history included?

Where the partner's vault kept it. PDM and PLM systems such as SolidWorks PDM, Windchill or Teamcenter hold released revisions, often the working versions between them, and the change orders and approvals behind each release. Some teams keep KiCad projects, which are text files, in git with commit-level history. Others kept only the latest released files, so the manifest states what history exists for each project.

Can a firm license designs it created for its customers?

Usually only with the customer's authorization. Design services firms, contract manufacturers and electronics manufacturing services providers often hold their customers' models, layouts and BOMs, but the designs belong to the customers under their agreements. Those files are excluded unless the customer agrees. Designs a firm developed for its own products are the usual basis for a dataset.

How are names and company details removed from CAD files?

Most identifying detail sits in metadata rather than geometry: title blocks, custom file properties, PDM user names and check-in comments, file paths and, on boards, silkscreen logos and part numbers. These are removed or replaced consistently across each project, since renaming files can break assembly references and library links. A recognizable product shape remains, and that is handled in the license terms.

Manufacturing quality and inspection records
Inspection results, nonconformances, dispositions and corrective actions from production plants
Proprietary codebases with full history
Complete private repositories with full version history, build files, tests and docs
Supply chain and logistics operations records
Linked orders, shipments, tracking events, exceptions and freight documents from real operations
Construction project records
RFIs, submittals, change orders, daily logs, inspections, photos and schedules from building projects
Field service and maintenance work orders
Work orders tracing symptom, diagnosis, parts and fix, with asset histories and photos

Evaluating this data for procurement?

Diligence packets are prepared per dataset. Rights, privacy processing and quality differ between datasets.

Request dataset diligence

Tell us what your models need

Send your spec — domain, volume, format, timeline and permitted use — and SourceX will match it against partner data and come back with what can be licensed.