HuggingAI4Engineering/cadgenbench-data

CADGenBench (Inputs)

Public inputs for the CADGenBench benchmark, which measures how well AI systems produce correct 3D mechanical parts as STEP files. This repository holds the task inputs only; the ground truth is withheld in a separate private repository so that the leaderboard's evaluation is the single source of truth.

• Leaderboard Space: HuggingAI4Engineering/CADGenBench
• Browse the tasks: the Tasks tab on the Space (thumbnails, search, generation/editing filter, per-task detail)
• Benchmark code: github.com/huggingface/cadgenbench
• Submissions + results: HuggingAI4Engineering/cadgenbench-submissions
• Ground truth: HuggingAI4Engineering/cadgenbench-data-gt (private)

Dataset Summary

CADGenBench contains 81 fixtures of mechanical parts split across two tasks:

• Generation (49): reproduce the part as an accurate 3D solid from an engineering drawing.
• Editing (32): apply a described change to an existing STEP solid.

Each submission is one output.step per fixture. Outputs are scored against the private ground truth by the CAD Score pipeline: a hard validity gate followed by a weighted mean of three orthogonal metrics (shape similarity, interface match, topology match). See the metric definitions.

Supported Tasks and Leaderboards

The benchmark is tool-agnostic: a submission is one STEP file per fixture, produced by any system (one LLM, several, a script, or by hand). Submit and view results on the leaderboard Space; the full submission contract (zip layout, meta.json, validity gate) is in docs/benchmark/submission.md.

Languages

Task prompts and descriptions are in English.

Dataset Structure

Data Instances

One directory per fixture, named by its numeric id. There are two shapes:

# Generation fixture
<id>/
├── description.yaml      # prompt + metadata
└── input.png            # the engineering drawing (input2.png, ... when multi-image)

# Editing fixture
<id>/
├── description.yaml      # prompt + metadata
├── edit_description.txt  # the requested change, as an instruction
├── input.step           # the starting solid to edit
├── input.mesh.npz       # trusted watertight mesh sidecar for input.step
└── renders/             # iso / front / top / right PNGs of the starting solid

Data Fields

description.yaml carries:

Field	Type	Description
description	string	The task prompt.
task_type	"generation" | "editing"	The task family.
input_files	list of strings	The input files the task declares (e.g. input.png, input2.png, or input.step).
input_type	"text+image" | "text+step"	Modality of the inputs.

Data Splits

No train/test split — all 81 fixtures form a single evaluation set (49 generation, 32 editing).

Dataset Creation

Source Data

The underlying CAD geometry is sourced from Mecado. Fixtures are real mechanical parts with mating interfaces (locating jigs, bolt patterns, slots). For each part, an engineering drawing (generation) or a starting solid plus an edit instruction (editing) forms the public input; the solved solid and its interface sub-volumes are held privately as ground truth.

Curation Rationale

The inputs are released publicly so contestants see exactly what they are solving, while the ground truth stays private so the leaderboard's server-side evaluation is the only path to a score.

Considerations for Using the Data

This is a benchmark input set, not a training corpus. To participate, generate one output.step per fixture and submit through the leaderboard Space.

Additional Information

Dataset Curators

The CADGenBench team (HuggingAI4Engineering).

Licensing Information

Released under the Open Data Commons Attribution License (ODC-BY).

Acknowledgements

CAD geometry sourced from Mecado. Thanks to the Mecado team.