---
license: mit
task_categories:
- image-to-text
tags:
- arxiv:2608.00799
- cad
- reverse-engineering
- mechanical-parts
- cadquery
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: part_id
    dtype: string
  - name: family
    dtype: string
  - name: mesh
    dtype: binary
  - name: mesh_file
    dtype: string
  - name: image
    dtype: image
  - name: prompt
    dtype: string
  splits:
  - name: train
    num_bytes: 13049973918
    num_examples: 3396
  download_size: 6400065980
  dataset_size: 13049973918
---

# CADENA-Bench

A benchmark for CAD reverse engineering on **real mechanical parts**, drawn from
MCB, TMCAD and CADNet, stripped of trivial and duplicate geometry, and grouped
into six part families so that a method's weaknesses are visible rather than
averaged away.

3396 parts. Reported per family, because aggregate scores hide exactly the
failure modes that matter: every learned method we evaluate loses roughly half
its GMS moving from sketch–extrude corpora to these parts, and the loss is
concentrated in revolved and patterned geometry.

- Paper: [CADENA: Stepwise CAD Reverse Engineering](https://arxiv.org/abs/2608.00799) (arXiv:2608.00799)
- Code: <https://github.com/zhemdi/cadena>
- Model: <https://huggingface.co/kulibinai/cadena>

## Families

| Family | Parts |
| --- | ---: |
| `shafts_and_bushings` | 762 |
| `gears_and_bearings` | 749 |
| `housings_and_frames` | 359 |
| `levers_and_profiles` | 463 |
| `springs_and_fasteners` | 1046 |
| `tooling_and_gauges` | 17 |
| **Total** | **3396** |

`tooling_and_gauges` is small enough that its column should not be read as a
ranking.

## Fields

| Column | Contents |
| --- | --- |
| `mesh_file` | Target mesh filename |
| `image` | The eight target views tiled into one image — six axis-aligned, two isometric |
| `ref` | Reference CAD program |
| `ref_path` | Source path of the reference program |
| `prompt` | Prompt used for stepwise inference |
| `family` | One of the six part families |

## Contamination

The 3504-part evaluation set contains 108 meshes byte-identical to a mesh in the
reinforcement-learning training corpus. Those are removed here, so **this
benchmark and the RL training set do not intersect**. The removal is
reproducible: hash every mesh in both sets and drop exact content matches.

## Evaluation

The paper scores predictions with GMS, which counts a predicted point as
matching a target point only when it is both close enough and carries a
consistent normal (`‖p − q‖ ≤ τ` and `n_p · n_q ≥ cos α`), so the score reflects
whether the surface was reconstructed rather than whether volume was filled.
A prediction that fails to build or is not watertight is invalid; invalid
predictions count toward the invalid rate and are excluded from the means.

```python
from datasets import load_dataset

ds = load_dataset("kulibinai/cadena-bench", split="train")
gears = ds.filter(lambda r: r["family"] == "gears_and_bearings")
```

## Citation

```bibtex
@article{cadena2026,
  title         = {CADENA: Stepwise CAD Reverse Engineering},
  author        = {Kabisov, Soslan and Savrasov, Gennadiy and Elistratov, Maksim and
                   Rodriguez, Antonio and Ignatiev, Daniil and Gavrilov, Nikita and
                   Uzdenov, Rustam and Boyko, Alexey I. and Pasechnik, Igor and
                   Konushin, Anton and Kuznetsov, Andrey and Zhemchuzhnikov, Dmitrii},
  year          = {2026},
  eprint        = {2608.00799},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2608.00799}
}
```