CAD Vision-Language Corpus · ADSKAILab

Zero-To-CAD-1m

Executable CAD construction sequences and multi-view renders for training parametric 3D computer-aided design models.

What Zero-To-CAD-1m is

Zero-to-CAD-1m is a dataset of 999,633 executable and interpretable computer-aided design (CAD) construction sequences developed by Autodesk Research. Synthesised entirely without real-world engineering files, the dataset pairs multi-view rendered images with CadQuery Python scripts that construct 3D boundary representation (B-Rep) solid geometry.

Prior parametric CAD datasets have largely been restricted to simple sketch-and-extrude sequences or lacked executable, human-readable construction code. Zero-to-CAD-1m addresses this by spanning 65 distinct part categories—such as structural brackets, pulleys, and housings—and incorporating a wide operational vocabulary including booleans, fillets, chamfers, shells, lofts, sweeps, and patterns. This provides computational researchers with validated parametric sequences for training vision-to-CAD and text-to-CAD models.

How it was made

The dataset was built using an agentic search framework where a large language model interacts in a multi-turn feedback loop with a CadQuery execution environment. The model is equipped with three tools: execute_and_validate to execute code and test geometric validity, lookup_documentation for TF-IDF retrieval over CadQuery API documentation, and grep_documentation for exact syntax matching. An initial protocol generates part descriptions across 65 categories, after which the agent iteratively writes, executes, and repairs the script until all validation criteria pass.

Each accepted sequence is required to produce a valid solid with at least seven B-Rep faces. The dataset records the CadQuery source code, extracted operation sequences, timing metrics, exported STEP and STL files, and eight 256×256 rendered perspectives (four front and four rear). In addition, precomputed 384-dimensional DINOv3 visual embeddings—averaged across the eight rendered views—are supplied alongside a FAISS IVF-PQ index for nearest-neighbour search.

Key facts

Total sequences: 999,633
Splits: Train (979,633), Validation (10,000), Test (10,000)
Part categories: 65
Minimum face count: 7 B-Rep faces
Available modalities: CadQuery Python code, 8 rendered images (256×256), STL mesh, STEP file
Visual embeddings: 384-dimensional DINOv3 (float32)
Zero-shot synthesis success rate: 22.3%
Mean attempts before success: 3.30

What you can use it for

Image-to-CAD reconstruction: Training vision-language models to reconstruct editable, parametric CadQuery code directly from multi-view 2D images.
Parametric CAD sequence generation: Training generative models to produce valid, sequential CAD construction operations and parameters from text or design specifications.
CAD program comprehension: Studying the semantic structure, operational ordering, and parameter dependencies of procedural 3D solid modelling scripts.
Generative model benchmarking: Evaluating the executable output, topological correctness, and complexity of generative CAD algorithms against validated geometric baselines.

Explore the data

The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.

Explore

Read the real rows without downloading anything

Reading rows…

LiveRead from ADSKAILab/Zero-To-CAD-1m at the moment you asked. Nothing is cached or stored — every row above came from that request.

How the columns are distributed

Over 7,370 rows of default/train

cadquery_ops_count

int

min: 3
median: 24
mean: 24.50
max: 80

cadquery_ops_json

string_text

min: 328
median: 3,812
mean: 3975.32
max: 14,213

face_latency_ms

float

min: 5.56
median: 38.49
mean: 68.65
max: 6740.13

num_faces

int

min: 6
median: 28
mean: 45.36
max: 5,415

num_renders

int

min: 8
median: 8
mean: 8
max: 8

ops_latency_ms

float

min: 0.13
median: 0.92
mean: 0.95
max: 11.22

uuid

string_text

min: 36
median: 36
mean: 36
max: 36

Strengths and limits

Good at

Every sample provides executable, human-readable CadQuery Python code featuring named parameters and logical construction steps.
Encompasses an extensive set of 3D operations beyond basic extrusion, including sweeps, lofts, boolean operations, fillets, chamfers, shells, and patterns.
Subjected to multi-stage geometric validation checking topological validity, contiguity, geometric complexity, and export fidelity.
Supplies multi-modal representations for every design, including multi-view renders, B-Rep STEP files, STL meshes, and precalculated visual embeddings.

Watch out for

Constructed entirely from synthetic generations, containing no real-world engineering files.
Language models can occasionally create designs that are locally plausible but lack global functional coherence.
Dimensions adhere to an arbitrary scale convention capped at 100 units rather than standardised engineering tolerances.
The distribution of CAD operations reflects language model priors rather than real-world industrial manufacturing frequencies.

Questions

Can I look inside Zero-To-CAD-1m without downloading it?

Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.

What licence is Zero-To-CAD-1m under?

Its authors declare apache-2.0. Read the terms on the original page before reusing the data.

What 3D modelling operations are represented in the dataset?

It contains sketch primitives (rectangles, circles, polygons, arcs, splines, slots), 3D operations (extrusion, cuts, revolutions, lofts, sweeps), modifications (fillets, chamfers, shells, offsets), booleans (union, cut, intersect), patterns (linear, polar, mirror), and features such as holes, threads, and ribs.

What file formats are provided for each CAD sample?

Each sample contains CadQuery Python source code, an exported STEP file (boundary representation), an exported STL file (triangular mesh), and eight 256×256 PNG image renders.

How are nearest neighbours retrieved across the collection?

A precomputed 1.54 GB file of 384-dimensional DINOv3 visual embeddings and a 97.7 MB FAISS IVF-PQ index are provided to perform approximate nearest-neighbour searches across all 999,633 samples.

What geometric validation steps were applied to the generated samples?

Generated scripts had to execute without errors in the CadQuery environment, maintain topological validity and contiguity, contain a minimum of seven B-Rep faces, and successfully export to STEP and STL formats.

How to cite it

Mohammadmehdi Ataei, Farzaneh Askari, Kamal Rahimi Malekshan, Pradeep Kumar Jayaraman. Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data. arXiv:2604.24479, 2026.

Written from Zero-To-CAD-1m's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.