CADGenBench Inputs: 81 Mechanical-Part Fixtures for AI CAD Generation from Engineering Drawings and STEP Editing (Hugging Face)

Provider
CADGenBench team (HuggingAI4Engineering), CAD geometry from Mecado
License
ODC-By-1.0 (source)
Size
360 files
Format
PNG, NumPy (in sampled files)
Stored at
Hugging Face (https://huggingface.co/datasets/HuggingAI4Engineering/cadgenbench-data)
Access
Free and public. No account with the data provider is needed.
Category
Code & Software Engineering
Source
https://github.com/huggingface/cadgenbench

About this dataset

Public task inputs for CADGenBench, a benchmark of how well AI systems produce correct 3D mechanical parts as STEP solids, run by the HuggingAI4Engineering team. It has 81 fixtures of real mechanical parts with mating interfaces such as locating jigs, bolt patterns and slots, sourced from Mecado. 49 generation fixtures ask a system to reproduce a part as an accurate 3D solid from an engineering drawing (one or more PNG images plus a text prompt); 32 editing fixtures give a starting STEP solid, a watertight mesh sidecar (input.mesh.npz), iso, front, top and right renders and a written instruction for a design change. Each fixture is a numbered folder with description.yaml (prompt, task type, input files, modality). Ground truth is held privately; submissions (one output.step per fixture) are scored on a leaderboard Space by a validity gate followed by shape similarity, interface match and topology match. About 360 files, 0.2 GB (measured 2026-11-21); open STEP files with CadQuery, OpenCascade or FreeCAD. This is an evaluation set, not a training corpus. Released under the Open Data Commons Attribution License (ODC-BY); see github.com/huggingface/cadgenbench.

How to load this dataset in Python

from huggingface_hub import hf_hub_download
path = hf_hub_download(repo_id="HuggingAI4Engineering/cadgenbench-data", filename="201/input.mesh.npz", repo_type="dataset")
print(path)

Data structure

File types in sampled files: PNG, NumPy

Files

Example file: 201/input.mesh.npz

Analyze this dataset without downloading it

On Scigantic, CADGenBench Inputs: 81 Mechanical-Part Fixtures for AI CAD Generation from Engineering Drawings and STEP Editing (Hugging Face) opens in a hosted JupyterLab notebook with the data already mounted, so you can read files where they are stored instead of copying them first. Each notebook includes an AI assistant that can see the mounted files and write and run analysis code. Open the dataset page to launch a notebook, or browse the full catalog.

This page is also available as Markdown and as JSON.

Related Code & Software Engineering datasets