Skip to content

Multimodal and embodied data

2D Engineering Drawings Paired with 3D CAD Models

Quick answer

A usable 2D drawing to 3D CAD dataset pairs each released drawing sheet with the exact model revision it documents, keyed by part number and revision, with views, dimensions, GD&T frames and notes kept intact. Public corpora give you models or construction sequences, but rarely the matching released drawings. Real pairs live in manufacturers' PDM and PLM vaults, so sourcing them means licensing operational engineering records, checking revision alignment, scrubbing title blocks and screening out export-controlled designs before anything moves.

By SourceX Editorial · Updated

  1. 01Your data spec
  2. 02Matched to businesses
  3. 03Rights and samples reviewed
  4. 04Licensed delivery
How it works with SourceX: Your data spec → Matched to businesses → Rights and samples reviewed → Licensed delivery.Illustrative

Why public CAD corpora rarely give you drawing-model pairs

Public CAD datasets mostly contain geometry without the released drawings a manufacturer actually signed off. The ABC dataset collected about a million CAD models for geometric deep learning, but it holds B-rep and mesh geometry, not dimensioned drawing sheets [1]. The Fusion 360 Gallery records sketch-and-extrude construction sequences for reconstruction, which is useful for sequence models but again carries no GD&T-annotated drawings [2].

Drawing benchmarks have the opposite gap. MechVQA is a mechanical-drawing question-answering benchmark built through a semi-automated construction and quality-control pipeline, not a conversion-to-solid dataset [3]. Text-to-CAD sets such as Text2CAD pair geometry with prompts rather than with drawings [4]; see text-to-CAD training data for that pairing.

The result is that a team converting a legacy archive of scanned D-size sheets into parametric models usually has to license pairs from companies whose engineering departments produced both artifacts under change control. That is the gap this page addresses; drawing OCR and annotation alone belong to engineering drawing annotation data.

What a training pair should contain

A training pair is the released drawing, the native or neutral model it documents, and a manifest that proves they describe the same part at the same revision. Specify each component explicitly, because suppliers will otherwise send whatever export is easiest.

  • Drawing side: the vector source where it exists (native drawing file, DWG/DXF or vector PDF) plus a raster at a stated DPI for scanned legacy sheets. Keep all sheets, including section views, detail views and the parts list.
  • Model side: the native part or assembly (SolidWorks, Creo, NX, Inventor, CATIA or Onshape exports) and a neutral STEP file, ideally AP242 so product manufacturing information can travel with the geometry. Ask whether the feature tree survives export or only the B-rep.
  • Annotations: dimensions, tolerances, datum feature symbols and feature control frames per ASME Y14.5 [5], surface finish symbols, weld symbols, general notes and the revision block. Views alone teach shape; annotations teach intent and are what consistency checkers need.
  • Manifest fields: part number, drawing number, drawing revision, model revision, PDM state (for example Released, Superseded, Obsolete), ECO or ECN number, CAD system and version, units, sheet size, projection angle (first or third angle) and whether the model was authored first or rebuilt from the drawing.

That last field matters more than most buyers expect. A model authored first and detailed into a drawing teaches forward documentation, while a model rebuilt from a legacy sheet by a drafter teaches exactly the conversion task you want to automate, including the drafter's interpretation choices.

Revision alignment is the main failure mode

The most common defect in drawing-model pairs is a revision mismatch, where the sheet documents one revision and the model file reflects another. In many shops, models are revised in PDM ahead of the drawing release, or a drawing is red-lined on paper and never fed back to the model; treat this as a working hypothesis to test in every sample, not as a known rate.

Concrete checks catch most of it. Compare the revision letter in the title block with the PDM revision of the model, confirm the ECO referenced in the revision table exists in the change history, and render the model in the drawing's projection to compare silhouettes against the drawing views. A dimension-level check, measuring the model at each dimensioned feature and comparing to the nominal on the sheet, catches subtler drift such as a hole moved in the model but not on the drawing.

Decide up front what you do with mismatches. For a conversion model, mismatched pairs are label noise and should be dropped or quarantined. For a drawing-model consistency checker, a labeled set of real mismatches with the ECO that caused them is valuable negative data, so ask suppliers to keep and flag them rather than silently remove them.

Illustrative example: invented to show structure; it does not describe an available dataset.

FieldExample valueWhy it matters
pair_idPAIR-000417Stable key across drawing, model and manifest
part_number[REDACTED-PN-0417]Join key; tokenized if it reveals the customer
drawing_rev / model_revC / CMust match for a clean conversion pair
pdm_stateReleasedExcludes WIP and unreleased checks
eco_refECO-[REDACTED]-2211Ties revision to change history
authoring_directionrebuilt_from_drawingSeparates forward detailing from legacy conversion
projectionthird_anglePrevents mirrored view interpretation
units / sheetinch / DNormalizes scale for raster inputs
gdt_presenttrueFilters for annotation-rich pairs
alignment_checksilhouette_pass; 2 of 38 dims out of toleranceFlags pairs for review
export_screenEAR99 per supplier reviewRecords classification before release

De-identifying title blocks and drawing content

Title blocks and notes carry the identities a supplier is least willing to release, so de-identification has to cover drawing text, model metadata and file names together. A typical title block names the supplying company, often the end customer, the drafter, checker and approver initials, a CAGE code, drawing and part numbers and sometimes a contract number.

Treat the drawing as an image and as text. Redact or replace title-block cells and customer-specific notes in the vector source, then regenerate the raster so no hidden layers or PDF annotation objects carry the original values. On the model side, scrub file properties, custom properties, author fields and assembly component names, and check STEP header fields such as FILE_NAME and author strings, which frequently preserve original paths and user names.

Keep part and drawing numbers as consistent tokens rather than deleting them, because the model needs the join key and your revision checks depend on it. Redaction itself changes the training distribution: a model trained only on blanked title blocks may learn that title blocks are empty. Ask for a consistent synthetic fill rather than black boxes if title-block reading is part of your task. The SourceX guide on de-identifying CAD and engineering drawings covers the method in more depth, and multimodal de-identification covers images and metadata in combination.

Export control and contract restrictions on engineering drawings

Engineering drawings are among the records most likely to be export-controlled, so every pair needs a classification decision before it leaves the supplier. As of October 2026, under ITAR, technical data includes information required for the design, development, production or manufacture of defense articles, including blueprints and drawings [6]. Under the EAR, releasing controlled technology to a foreign person inside the United States can count as an export to that person's country, which matters if your annotation or ML team includes foreign nationals [7].

Ask suppliers to exclude drawings flagged ITAR or with an ECCN other than EAR99 unless your own export counsel has approved the handling. Also exclude designs made under customer contracts that restrict reuse of drawings, and drawings marked proprietary to a customer rather than the supplier. Distribution statements, CUI markings and "property of" legends on the sheet are strong signals to pull a record. The export-controlled technical data guide lays out the ITAR and EAR checks in more detail.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Choosing pairs for conversion, consistency checking and copilots

Different applications need different slices of the same archive, so state your task before you request data. A pair that is ideal for one model can be noise for another.

ApplicationPairs to prioritizePairs to exclude or label
Legacy drawing-to-CAD conversionScanned or vector legacy sheets with models rebuilt from them; varied drafting standards and erasDrawings auto-generated from the model with no human edits
Drawing-model consistency checkingClean released pairs plus labeled real mismatches with ECO contextUnlabeled mismatches
CAD copilot (detailing assistance)Model-first pairs with full GD&T and notes; native files with feature treesRaster-only drawings without vector source
Evaluation setHeld-out parts from a different supplier or product line, with independent revision verificationAny part family present in training

Diversity across suppliers matters more than raw volume for conversion work. Drafting conventions vary by company, era and CAD system: first versus third angle projection, pre-Y14.5-2018 symbology, hand-lettered notes and stamped revisions on older sheets. A model trained on one company's template will often fail on another's title block and view layout. For held-out sets, see private multimodal evaluation sets.

Request template for a drawing-to-CAD dataset

A good request describes the data, the alignment rules and the exclusions, not the companies you hope will hold it. You can adapt the multimodal dataset specification template or start from this outline.

Illustrative example: invented to show structure; it does not describe an available dataset.

  1. Task: convert scanned 2D mechanical part drawings into parametric 3D models; secondary task flags drawing-model inconsistencies.
  2. Domain: machined and sheet-metal parts; assemblies out of scope initially.
  3. Drawing format: vector PDF or DWG where available; scans at a stated minimum DPI; all sheets.
  4. Model format: native files plus STEP AP242; feature tree preferred.
  5. Alignment: part number and revision must match; mismatches retained only with a flag and ECO reference.
  6. Annotations: GD&T per Y14.5 retained; notes and revision tables retained.
  7. De-identification: title-block names, customer references, CAGE codes and file metadata replaced with consistent tokens.
  8. Exclusions: ITAR, non-EAR99 ECCNs, customer-owned designs, CUI or distribution-limited drawings.
  9. Allowed use: training and evaluation of conversion and checking models; state whether derivative weights may be deployed commercially.
  10. Sample: a small set of pairs for alignment and redaction review before any larger scope.

Licensing these records can involve more than one rightsholder, for example a supplier that drew a part to its customer's design. The guide to licensing multimodal records from several rightsholders covers how to ask who owns each side of the pair.

How SourceX approaches drawing and CAD pairs

SourceX sources operational datasets from US companies, including engineering records and documents, and manages the licensing and ongoing purchases. Drawing and model pairs are sourced on request rather than held in stock, and a request does not guarantee a match. Buyers describe the data; SourceX looks for US businesses that hold it, and every release is approved by the supplying company.

Each dataset is rights-reviewed for ownership and consents and delivered under a license that defines the records, uses, term and delivery. Personal details such as names and emails are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect. Delivery runs through private, access-controlled workflows only after an executed agreement and supplier approval. You can describe the drawing and model pairs you need at any stage of scoping. Related owner pages cover CAD and PCB design datasets and licensing CAD and engineering drawings, and the multimodal data hub maps adjacent pairings.

Source drawing-to-CAD training pairs

If you are building drawing-to-CAD conversion or consistency checking, send SourceX a description of the parts, formats, alignment rules and exclusions you need. SourceX assesses data and licensing permissions with suppliers, and nothing is contracted until a supplier agrees. Start a buyer request.

Sources

  1. Koch et al., arXiv (CVPR 2019), "ABC: A Big CAD Model Dataset For Geometric Deep Learning" (2019). https://arxiv.org/pdf/1812.06216
  2. Willis et al., arXiv, "Fusion 360 Gallery: A Dataset and Environment for Programmatic CAD Reconstruction" (2020). https://arxiv.org/abs/2010.02392v2
  3. arXiv, "MechVQA: Benchmarking and Enhancing Multimodal LLMs on Comprehensive Mechanical Drawing Understanding" (2026). https://arxiv.org/pdf/2605.30794
  4. arXiv, "Text2CAD: Generating Sequential CAD Models from Beginner-to-Expert Level Text Prompts" (2024). https://arxiv.org/pdf/2409.17106
  5. ASME, "Y14.5 - Dimensioning and Tolerancing". https://qa12.asme.org/codes-standards/find-codes-standards/y14-5-dimensioning-tolerancing
  6. eCFR (U.S. Government Publishing Office), "22 CFR 120.33 -- Technical data". https://www.ecfr.gov/current/title-22/chapter-I/subchapter-M/part-120/subpart-C/section-120.33
  7. Bureau of Industry and Security, "Deemed exports". https://www.bis.gov/deemed-exports

Tell us what your models need

Share scope, volume, language, format, timing and licensing requirements.

Request data