---
license: mit
pretty_name: CAD-bench Task Payloads
language:
- en
tags:
- cad
- benchmark
- build123d
- step
- mlcroissant
- language-model-evaluation
size_categories:
- n<1K
---

# CAD-bench Task Payloads

This dataset contains the public task payloads for CAD-bench, an executable
benchmark for language-model CAD agents. Each task directory includes:

- `prompt.txt`: the natural-language benchmark prompt
- `task.toml`: task metadata, difficulty, evaluator name, and expected values
- `gold.py`: a reference Build123D solution used for validation and media generation
- optional fixtures such as STEP files or Blender simulation scripts

The benchmark runtime and scoring code are distributed with the accompanying
anonymous submission artifact. The runtime loads this dataset by setting:

```bash
HF_TASKS_REPO_ID=uhiguys/cad-bench-ed-2026-anonymous
HF_TASKS_REVISION=0ef63c7666d04fcc7e8061e23a37b4ebce75c4a7
```

## Intended Use

Use this dataset with the CAD-bench runtime to evaluate CAD code-generation or
agentic CAD systems. Scores are diagnostic benchmark signals; they are not
certifications that generated mechanical parts are safe to manufacture or deploy.

## Data Provenance

The tasks are synthetic CAD prompts and benchmark metadata authored for this
benchmark. They do not contain personal data or human-subject records. The M3
socket-head task includes a STEP fixture for a standard commercial fastener,
documented in the task metadata and paper license table.

## Limitations

The release has 17 tasks. Some simple geometry tasks are close to solved by
current models, while functional assembly tasks remain difficult. The current
reference implementations use Build123D, although the benchmark is intended to
score submitted CAD artifacts rather than a particular modeling API.

## Citation

Anonymous Author(s). CAD-bench: An Executable Benchmark for Language-Model CAD
Agents. NeurIPS Evaluations & Datasets submission, 2026.