OmniMech — 5K Dataset and Evaluation Code

This release contains 5,000 CAD samples and code for SOLIDWORKS conversion, CAD rendering, drawing-to-CAD model queries, and geometric evaluation. Each sample includes a native parametric part, STEP geometry, an STL mesh, an engineering drawing, rendered views, a video, and parametric feature data with Python reconstruction code.

Package structure

OmniMech_dataset_release_5k/
├── README.md
├── dataset/
│   ├── manifest.json
│   ├── manifest.csv
│   ├── 000001/
│   ├── 000002/
│   └── … 005000/
├── preview/
│   └── all-00000-of-00001.parquet
└── evaluation/
    ├── README.md
    ├── dependencies/
    │   ├── environment.yml
    │   ├── requirements.txt
    │   └── requirement_agentic.txt
    ├── evaluation/
    │   ├── align_stl_icp_chamfer.py
    │   └── align_stl_icp_eval_improved.py
    ├── rendering/
    │   ├── render_cad_wireframe.py
    │   ├── render_cad_wireframe_cube_vertices.py
    │   ├── render_cad_wireframe_all_angle.py
    │   └── render_cad_walkthrough_video.py
    ├── solidworks_conversion/
    │   └── convert_solidworks_to_stl_step.py
    └── llm_query/
        ├── zeroshot/
        └── agentic/

See evaluation/README.md for environment setup, SOLIDWORKS conversion, single-view/multi-view/video rendering, model queries, and geometric evaluation. Native SOLIDWORKS conversion requires Windows, an installed SOLIDWORKS application, and pywin32.

Dataset Viewer

The preview/ Parquet table contains one row for each of the 5,000 samples. Its drawing and render columns display compact image previews. Path columns point to the corresponding files under dataset/, relative to the repository root. The all split includes the complete release; it is not a training or evaluation partition.

Dataset structure

Samples use six-digit identifiers from 000001 to 005000. The identifier is shared by the sample folder and its files.

dataset/000001/
├── metadata.json
├── raw/000001.sldprt
├── brep/000001.step
├── mesh/000001.stl
├── drawings/000001.pdf
├── drawing_bitmap/000001.png
├── 8_view_color_brep/000001_view_01.png … 000001_view_08.png
├── 8_view_plain_brep/000001_view_01.png … 000001_view_08.png
├── 20_view_color_brep/000001_view_01.png … 000001_view_20.png
├── 20_view_plain_brep/000001_view_01.png … 000001_view_20.png
├── single_view_color_brep/000001.png
├── single_view_plain_brep/000001.png
├── video_brep/000001.mp4
└── parametric_code/
    ├── 000001.json
    └── 000001.py

Folder or file	Contents per sample
raw/	Native SOLIDWORKS part (.sldprt).
brep/	Boundary representation in STEP format (.step).
mesh/	Triangle mesh in STL format (.stl).
drawings/	One engineering drawing PDF.
drawing_bitmap/	One PNG of the engineering drawing.
8_view_color_brep/, 8_view_plain_brep/	Eight PNG renders in each folder.
20_view_color_brep/, 20_view_plain_brep/	Twenty PNG renders in each folder.
single_view_color_brep/, single_view_plain_brep/	One PNG render in each folder.
video_brep/	One MP4 visualization.
metadata.json	Sample identifier, asset paths, drawing information, and parametric-code details.
parametric_code/	Feature-model JSON and matching Python reconstruction script.

View numbers are local to each view folder. Color and plain sets share the same filename pattern.

Finding samples

The two manifests are inside dataset/:

• manifest.json contains the dataset summary and one record per sample. General assets are listed under assets; parametric-code paths and conversion details are grouped under parametric_code.
• manifest.csv is a compact table with one row per sample, including identifiers, drawing-page count, asset paths, and Python completeness information.

All manifest paths are relative to the dataset/ directory containing the manifests, for example 000001/mesh/000001.stl. Paths inside a sample's metadata.json are relative to that sample folder. Read CSV sample_id values as strings to retain leading zeros. Python, parametric-code, and JSON metadata fields appear after the general asset information.

For example, sample 000001 has its reference mesh at dataset/000001/mesh/000001.stl and its drawing image at dataset/000001/drawing_bitmap/000001.png.

Parametric data and Python completeness

The JSON in parametric_code/ contains the extracted feature tree, modeling features, dimensions, units, and other available feature data. The matching Python script uses CadQuery and NumPy and reads the adjacent JSON file by default. Keep the pair together when moving a sample.

Where present, data_complete: false or feature_data_complete: false indicates incomplete feature data.

Python completeness	Samples	Meaning
full	1,871	Export status ready; the exporter reports no unsupported feature types.
partial	3,129	The Python reconstruction has unsupported features or other exporter limitations. The JSON retains extracted data for unsupported features where available.

These labels describe conversion coverage. They do not guarantee successful Python execution or geometric equivalence to the native part.