CAD Vision-Language Corpus · ADSKAILab
Zero-To-CAD-100k
A curated 100,000-sample dataset of diverse parametric CAD programs and multi-view renders for 3D model generation.
What Zero-To-CAD-100k is
Zero-To-CAD-100k is a curated vision-language dataset of 100,000 computer-aided design (CAD) construction sequences developed by researchers at Autodesk Research. It is a diverse subset of the full one-million-sample Zero-to-CAD collection, built via agentic synthesis to provide interpretable CAD programs without relying on real-world engineering proprietary data.
Training machine learning systems to infer CAD models directly from images or text often requires large amounts of paired visual and parametric data. This dataset addresses compute constraints associated with training and evaluating on million-scale corpuses by offering a representative, geometrically diverse benchmark tailored for lightweight fine-tuning, rapid experimentation, and evaluation.
Each entry in the dataset pairs executable Python source code written for the CadQuery library with multi-view 2D rendered images, standard boundary representation (B-Rep) STEP files, and triangulated STL meshes. This multimodal structure enables computational researchers to bridge visual understanding and procedural solid-modelling code.
How it was made
The 100,000 samples were selected from the full 1M dataset through an embedding and clustering pipeline designed to maximise geometric diversity. Every 3D model was first rendered across eight viewpoints and encoded using visual features extracted from DINOv3, which were averaged across views into a single visual embedding.
K-means clustering was then performed across the visual embedding space to establish 100,000 distinct clusters. The exemplar closest to each cluster centroid was selected for inclusion, ensuring the final corpus spans the broad distribution of part types, geometric complexity, and construction operations present in the source dataset.
Key facts
- Total samples
- 100,000
- Split counts
- Train: 81,015; Validation: 9,734; Test: 9,767
- Primary modalities
- Executable CadQuery Python code, 256×256 rendered images, STL meshes, STEP solid models
- Data format
- Parquet
- Rendered views per sample
- 8
- Sketch primitives
- rect, circle, polygon, arc, spline, slot
- 3D modelling operations
- extrude, cut, revolve, loft, sweep, fillet, chamfer, shell, offset, union, intersect, linear/polar/mirror patterns, holes, threads, ribs
What you can use it for
- Benchmarking Image-to-CAD Models
- Evaluate the accuracy and reconstructive performance of vision-language models generating parametric CAD code from multi-view images using standardised test splits.
- Lightweight Model Fine-Tuning
- Train or fine-tune compact vision-language models, such as Qwen3-VL-2B, on parametric code generation without the compute overhead of million-scale datasets.
- Procedural Geometry and CAD Synthesis
- Study generative algorithms that output executable CadQuery Python scripts containing boundary representation solids and geometric operations.
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-100k at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 9,298 rows of default/train
cadquery_ops_count
int- min
- 3
- median
- 23
- mean
- 23.75
- max
- 82
cadquery_ops_json
string_text- min
- 317
- median
- 3,725
- mean
- 3840.69
- max
- 21,235
face_latency_ms
float- min
- 6.03
- median
- 26.15
- mean
- 39.09
- max
- 1793.49
num_faces
int- min
- 2
- median
- 25
- mean
- 34.66
- max
- 1,446
num_renders
int- min
- 8
- median
- 8
- mean
- 8
- max
- 8
ops_latency_ms
float- min
- 0.13
- median
- 0.71
- mean
- 0.77
- max
- 3.31
uuid
string_text- min
- 36
- median
- 36
- mean
- 36
- max
- 36
Strengths and limits
Good at
- Selected through centroid-based clustering over DINOv3 embeddings to maximise geometric variety across the 100,000 samples.
- Includes executable CadQuery Python code alongside boundary representation (STEP) and surface mesh (STL) formats for each record.
- Provides eight canonical 256×256 image views per model to support vision-to-geometry learning tasks.
- Pre-split into explicit train, validation, and test subsets to facilitate reproducible benchmarking.
Watch out for
- Restricted to a 100,000-example sample, meaning it lacks the maximum data volume of the parent 1M corpus for large-scale pre-training.
- Contains synthetic CAD sequences rather than designs extracted directly from human engineering workflows.
- Image renders are fixed to eight views at a 256×256 resolution.
Questions
Can I look inside Zero-To-CAD-100k 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-100k under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
What 3D formats are included with each sample?
Each sample contains executable CadQuery Python code, exported STEP solid boundary-representation files, triangulated STL mesh files, and eight 256×256 rendered PNG images.
How were the 100,000 models chosen from the 1M dataset?
The full dataset was encoded by averaging 8-view DINOv3 visual embeddings, grouped into 100,000 clusters via K-means, and the nearest sample to each centroid was extracted.
How are the training, validation, and test splits divided?
The dataset is divided into 81,015 training examples, 9,734 validation examples, and 9,767 testing examples.
Can the CAD scripts in this dataset be run directly in Python?
Yes, each sample contains a CadQuery Python script that can be executed in an environment with the CadQuery library installed to reproduce the 3D solid geometry.
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-100k'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.