MechIntelligence/manufacturing-digital-gene-dataset

Manufacturing Digital Gene Dataset

Overview

The Manufacturing Digital Gene Dataset is a large-scale dataset designed for geometry-aware manufacturing intelligence and CAD generation. It provides structured digital representations of CAD models by combining executable CAD programs, manufacturing-oriented digital genes, metadata, and 3D geometry.

The dataset is intended to support research in:

• CAD code generation
• Text-to-CAD
• Image-to-CAD
• Geometry understanding
• Manufacturing knowledge representation
• Embodied AI
• Robotic assembly and disassembly
• Digital Twin
• Foundation Models for Mechanical Design

Statistics

Split	Samples
Train	32,000
Validation	4,000
Test	4,000
Total	40,000

The dataset is released as Parquet shards for efficient loading and scalable training.

Directory Structure

manufacturing-digital-gene-dataset/
├── README.md
├── train/
│   ├── train-00000.parquet
│   ├── train-00001.parquet
│   └── ...
├── val/
│   ├── val-00000.parquet
│   └── ...
└── test/
    ├── test-00000.parquet
    └── ...

Sample Fields

Each row in the Parquet dataset contains:

Field	Description
sample_id	Unique sample identifier
split	Dataset split, including train, val, or test
cadquery_code	Executable CADQuery program
metadata_json	Metadata describing the CAD model
digital_gene_json	Manufacturing Digital Gene representation
step_bytes	STEP CAD model stored as binary
stl_bytes	STL mesh stored as binary
renders	Rendered multi-view images stored as binary

Digital Gene Representation

Each sample includes a structured manufacturing digital gene describing:

• Geometric primitives
• Feature parameters
• Manufacturing semantics
• Constraint information
• Topological relationships
• Design attributes

The Digital Gene representation serves as a compact, machine-readable description of manufacturing knowledge and can be used as an intermediate representation for CAD generation and downstream reasoning.

Data Loading Example

from datasets import load_dataset

dataset = load_dataset(
    "sijiahui/manufacturing-digital-gene-dataset",
    data_files={
        "train": "train/*.parquet",
        "validation": "val/*.parquet",
        "test": "test/*.parquet",
    }
)

print(dataset)
print(dataset["train"][0].keys())

Potential Applications

• CAD Foundation Models
• CAD Program Generation
• Text-to-CAD
• Image-to-CAD
• Manufacturing Intelligence
• Mechanical Design Automation
• Digital Twin
• Robotic Manipulation
• Human-Robot Collaboration
• Geometry-aware Large Language Models

Acknowledgements

We would like to acknowledge ADSKAILab/Zero-To-CAD-100k for providing an important reference and foundation for CAD-oriented generative modeling research.

The Manufacturing Digital Gene Dataset is developed with inspiration from Zero-To-CAD-100k, and further extends the dataset representation toward manufacturing-oriented digital genes, structured CAD semantics, geometry-aware reasoning, and downstream applications in embodied AI, robotic assembly/disassembly, and digital twin systems.

License

This dataset is released under the CC BY-NC 4.0 license unless otherwise specified.

Contact

For questions, suggestions, or collaborations, please open an issue on the Hugging Face repository or contact the authors.