---
task_categories:
- image-to-text
dataset_info:
  features:
  - name: file_id
    dtype: string
  - name: label
    dtype: string
  - name: step
    dtype: binary
  - name: stl
    dtype: binary
  - name: obj
    dtype: binary
  - name: glb
    dtype: binary
  - name: singleview_image
    dtype: image
  - name: multiview_image
    dtype: image
  - name: pbr
    dtype: image
  - name: noisy_stl
    dtype: binary
  splits:
  - name: bench0
    num_bytes: 8687768922
    num_examples: 3000
  - name: bench0F
    num_bytes: 9820522555
    num_examples: 3000
  - name: bench1A
    num_bytes: 14881149982
    num_examples: 3000
  - name: bench1B
    num_bytes: 15182316852
    num_examples: 3000
  - name: bench2
    num_bytes: 9795102660
    num_examples: 3000
  - name: bench3
    num_bytes: 43616021783
    num_examples: 3000
  download_size: 80229154806
  dataset_size: 103734295182
configs:
- config_name: default
  data_files:
  - split: bench0
    path: data/bench0-*
  - split: bench1A
    path: data/bench1A-*
  - split: bench1B
    path: data/bench1B-*
  - split: bench2
    path: data/bench2-*
  - split: bench3
    path: data/bench3-*
  - split: bench0F
    path: data/bench0F-*
---

# CADBench

CADBench is a unified multimodal benchmark for AI-assisted CAD program generation, introduced in the paper [CADBench: A Multimodal Benchmark for AI-Assisted CAD Program Generation](https://huggingface.co/papers/2605.10873).

The benchmark contains 18,000 evaluation samples spanning six benchmark families derived from DeepCAD, Fusion 360, ABC, MCB, and Objaverse. It is designed to measure progress in editable 3D reconstruction and multimodal CAD understanding.

### Dataset Summary

CADBench supports evaluation across five input modalities:
- **Clean meshes**
- **Noisy meshes**
- **Single-view renders**
- **Photorealistic renders (PBR)**
- **Multi-view renders**

The benchmark evaluates models across six metrics covering geometric fidelity, executability, and program compactness. STEP-based families are stratified by B-rep face count to support controlled analysis across complexity and object variation.