MUSE: A CAD Design Benchmark with Multi-modal Ground Truth and Rubric-based Evaluation

MUSE is a benchmark of 106 CAD design cases for evaluating language and multi-modal models on engineering-grade 3D design tasks. Each case pairs a natural-language design specification with multi-view ground-truth artefacts (2D engineering drawings + 3D rendered images) and a hand-crafted, rubric-style evaluation guide.

Why this benchmark

Most CAD/3D benchmarks evaluate either pure geometry reconstruction or free-form text-to-3D quality, but not both at once and not with the engineering correctness lens (assembly topology, joint design, tolerance, function). MUSE fills this gap by:

1. providing a textual design specification as input,
2. offering two complementary ground-truth views — a four-view dimensioned engineering drawing and a 3D STP render — so judges can reason about both geometry and manufacturability,
3. shipping a per-case rubric with explicit pass/fail criteria for assembly readiness, joint design, tolerance, functional adaptation, and more.

Dataset structure

.
├── README.md
├── LICENSE
├── metadata.jsonl                # one JSON object per case
└── cases/
    └── <case_id>/
        ├── design_description.md     # the input prompt
        ├── <case_id>.png             # 4-view dimensioned engineering drawing (from SVG)
        ├── <case_id>_stp_render.png  # 3D STP rendered image
        └── evaluation_rubric.md      # rubric for judging a generated design

Each line of metadata.jsonl looks like:

{
  "case_id": "bookshelf",
  "design_description": "cases/bookshelf/design_description.md",
  "svg_png":           "cases/bookshelf/bookshelf.png",
  "stp_render":        "cases/bookshelf/bookshelf_stp_render.png",
  "evaluation_rubric": "cases/bookshelf/evaluation_rubric.md"
}

For 9 organic / curved-vase cases (vase_wave_*, wave_vase), where 2D SVG projection cannot capture the surface geometry, the 3D STP render image is intentionally identical to the engineering-drawing PNG — i.e. the curved silhouette image is used as both ground-truth views.

How to use

from datasets import load_dataset

ds = load_dataset("dongxiaoyu/MUSE", split="test")
print(ds[0])
# {
#   'case_id': 'bookshelf',
#   'design_description': 'cases/bookshelf/design_description.md',
#   ...
# }

To resolve a referenced file:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="dongxiaoyu/MUSE",
    repo_type="dataset",
    filename=ds[0]["design_description"],
)
print(open(path).read())

Recommended evaluation protocol

For each case:

1. Feed design_description.md to the model under test as the user prompt.
2. Collect the model's design output (e.g. CAD code, SVG, mesh, or rendered image, depending on the modality being evaluated).
3. Compare the model output against the ground-truth pair (<case>.png, <case>_stp_render.png).
4. Apply each criterion in evaluation_rubric.md as an independent 0/1 score and report per-rubric pass-rates as well as the overall mean.

A reference judge prompt and breakdown JSON schema are included in the companion paper.

Cases

The benchmark covers 106 design cases across the following families: chairs, tables, bookshelves & shelves, CNC furniture (TV stands, shoe racks, desks), pegboards, pen holders, stools, vases, and miscellaneous household fixtures (handle, hook, toothbrush holder, etc.).

Responsible AI metadata

A complete Croissant 1.0 metadata file with the MLCommons RAI extension is shipped at the root of the repository as croissant.json (raw URL: croissant.json). The fields below summarize the RAI sections required by the NeurIPS 2026 Datasets & Benchmarks track.

Uses (intended and out-of-scope)

Intended use. Benchmarking large language models and vision-language models on (a) interpreting natural-language CAD design specifications, (b) generating or reasoning about engineering CAD geometry, and (c) reasoning about manufacturability concepts including assembly topology, joint type selection, tolerance, and functional adaptation. The 0/1 rubric score, aggregated across criteria and cases, is meant to be reported alongside qualitative inspection of model outputs.

Out-of-scope use. This dataset is not designed as a training corpus. It contains only 106 cases and would not provide useful coverage as pre-training data; using it for training risks contamination of any benchmark report computed on the same cases. The benchmark also does not validate physical fabricability — passing every rubric is a necessary but not sufficient condition for a design to be machinable in the real world.

Bias, risks, and limitations

Biases. Domain coverage is biased toward Western-style household objects and small fixtures (chairs, tables, shelves, vases, holders, pegboards, stools); automotive, aerospace, electronics, and large civil-engineering domains are absent. Manufacturing methods are limited to CNC milling, 3D printing, traditional joinery, and a few adhesive/fastening methods (no casting, forging, or metal additive). Materials are dominated by timber, plastic, and metal. Component counts skew toward 10–60 parts. All design specifications are in English. Because geometry seeds were generated with LLM assistance, stylistic or naming biases of the underlying foundation models may propagate into the cases.

Limitations. Only 4-view orthographic + isometric drawings and a single 3D render are provided as ground truth — there is no live STEP-vs-STEP geometric comparison built into the rubric. STEP files are watertight but tolerances and clearances are not validated against physical fabrication. Rubric scoring is binary (Pass / Fail) per criterion, so mostly-correct designs are not credited proportionally. Each rubric is authored against one specific reference design and may be brittle when applied to substantially different but equally valid alternative designs. With 106 cases the benchmark cannot finely separate models with very small performance differences.

Personal and sensitive information

The dataset contains no personally identifiable information. There are no human-subject data, no images of people, no audio or video, no health, financial, geographic-location, biometric, or genetic data. None of the categories tracked by the Croissant RAI specification appear in the dataset.

Source data

The benchmark is fully original. CAD geometry seeds were generated with assistance from leading commercial large language models (GPT-4 family, Claude, Gemini) between 2026-04 and 2026-05, executed in a sandboxed CadQuery environment, then curated and verified by the authors. The natural-language design specifications and the per-case evaluation rubrics were authored entirely by the human authors. No third-party copyrighted CAD models, scraped web content, or derived datasets are included.

Synthetic data

All CAD geometry, all design descriptions, and all rubrics are synthetic or human-authored. There is no real-world telemetry, sensor capture, or photographic content in the dataset.

Social impact

Positive. Rigorous evaluation of LLM/VLM ability on engineering tasks supports research into AI-assisted design tools that augment human designers, lower the barrier to entry for novice mechanical designers, and enable faster iteration on early-stage product design.

Negative. Improved model capabilities on engineering-grade CAD generation could be misused to reproduce proprietary mechanical designs without authorization or to bypass professional review in safety-critical applications. The dataset itself contains no operational hazards (no weapons, surveillance content, or dual-use components).

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to share and adapt the material for any purpose, including commercially, provided that you give appropriate credit.

Citation

@inproceedings{muse_2026,
  title     = {MUSE: A CAD Design Benchmark with Multi-modal Ground Truth and Rubric-based Evaluation},
  author    = {Dong, Xiaoyu and others},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track},
  year      = {2026}
}