---
license: cc-by-4.0
task_categories:
- visual-question-answering
language:
- en
- de
tags:
- engineering
- drawing
- CAD
pretty_name: Technical drawings for Manufacturability Benchmark
size_categories:
- n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: "techmb.tsv"
---


# Dataset Card for TechMB
## Dataset Details

The Technical drawing for Manufacturability Benchmark (TechMB) gives a domain specific benchmark for the task of manufacturability evaluations based on technical drawings.
This task is described as a Visual Question Answering (VQA) task targeted at Vision Language Models (VLM) consisting of 947 question-answer pairs on 180 distinct techical drawings.
The objects, the technical drawings are developed from, represent a selection of parts of the [Fusion 360 Gallery Segmentation Dataset](https://github.com/AutodeskAILab/Fusion360GalleryDataset/tree/master). 
Please refer to [their publication](https://doi.org/10.48550/arXiv.2104.00706) for further information. Their licence statement can be found [here](https://github.com/AutodeskAILab/Fusion360GalleryDataset/blob/master/LICENSE.md).
The IDs of the parts from the f360 segmentation dataset also declare the corresponding technical drawings for better association.

- **Curated by:** Leonhard Kunz
- **Funded by:** Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - Projektnummer (543073350)
- **Language(s) (NLP):** English, German
- **License:** CC-BY-4.0


## Dataset Structure

The dataset consists contains the following fields:

- **task_id:** ID of the specific question.
- **eval_type:** Classifier for the expected answer type (answer matching or multiple choice).
- **drw_id:** ID of the part and the corresponding drawing.
- **image:** Bit64 encoded image of the exported technical drawing.
- **drw_complexity:** Numeric complexity of the drawing. Calculated with the following formula: $complexity=(faces+dimensionings+\frac{annotation characters}{4.6})*views$
- **question:** The question text.
- **answer:** The expected answer corresponding to the answer type.
- **label_confidence:** The confidence of the assorted labels in manual labelling (low, medium, high).



## Citation:
Please refer to our dataset using the following DOI: [doi:10.57967/hf/6214](https://doi.org/10.57967/hf/6214)

For more information, refer to our publication:
```
@inproceedings{kunz2025techmb,
  title={TechMB: Exploring the Potential of Vision Language Models for Interpreting Technical Drawings},
  author={Kunz, Leonhard and Klostermeier, Mario and Thanabalan, Kokulan and Legler, Tatjana and Ruskowski, Martin and others},
  booktitle={DS 140: Proceedings of the 36th Symposium Design for X (DFX2025)},
  pages={1--10},
  year={2025},
  doi={https://doi.org/10.35199/dfx2025.19}
}
```