README.md

---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: bpmn
    dtype: string
  - name: image_filename
    dtype: string
  - name: bpmn_filename
    dtype: string
  - name: split
    dtype: string
  splits:
  - name: train
    num_bytes: 43277266.0
    num_examples: 101
  - name: validation
    num_bytes: 110828879.0
    num_examples: 50
  - name: test
    num_bytes: 104260192.0
    num_examples: 51
  download_size: 220317996
  dataset_size: 258366337.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
task_categories:
- image-to-text
- image-text-to-text
pretty_name: BPMN Diagram ↔ BPMN XML Paired Dataset
size_categories:
- 1K<n<10K
---

# 🏗️ BPMN Diagram → BPMN XML Paired Dataset  
### Structured Extraction from Business Process Diagrams using Vision-Language Models

This dataset contains **Business Process Model and Notation (BPMN)** diagrams paired with their corresponding `.bpmn` XML ground truth files.  
The dataset is designed for **training, evaluation, and benchmarking** multimodal models that perform *structured extraction from diagrams*, including OCR-enhanced pipelines and vision-language models (VLMs).

---

# 📦 Dataset Contents

Each example includes:

| Field | Description |
|-------|-------------|
| `image` | BPMN diagram image (PNG/JPEG), uploaded directly to HF |
| `bpmn` | Text content of the corresponding `.bpmn` XML file |
| `image_filename` | Original image filename |
| `bpmn_filename` | Original BPMN filename |
| `split` | One of: `train`, `validation`, `test` |

Folder structure used during creation:

```
dataset/
├── train/
│   ├── images/
│   ├── bpmn/
├── validation/
│   ├── images/
│   ├── bpmn/
├── test/
    ├── images/
    ├── bpmn/
```

---

# 🖼️ Example Image

![Example BPMN Image](https://huggingface.co/datasets/pritamdeka/BPMN-VLM/resolve/main/process_001.png)


### Ground Truth BPMN (excerpt)
```<?xml version='1.0' encoding='UTF-8'?>
<bpmn:definitions xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
    xmlns:bpmn="http://www.omg.org/spec/BPMN/20100524/MODEL"
    xmlns:bpmndi="http://www.omg.org/spec/BPMN/20100524/DI"
    xmlns:dc="http://www.omg.org/spec/DD/20100524/DC"
    id="Definitions_1"
    targetNamespace="http://bpmn.io/schema/bpmn">

  <bpmn:process id="Process_1" isExecutable="false">

    <bpmn:startEvent id="StartEvent_1">
      <bpmn:outgoing>Flow_0abcd12</bpmn:outgoing>
    </bpmn:startEvent>

    <bpmn:task id="Task_1" name="Receive Order">
      <bpmn:incoming>Flow_0abcd12</bpmn:incoming>
      <bpmn:outgoing>Flow_0efgh34</bpmn:outgoing>
    </bpmn:task>

    <bpmn:exclusiveGateway id="Gateway_1">
      <bpmn:incoming>Flow_0efgh34</bpmn:incoming>
      <bpmn:outgoing>Flow_0ijkl56</bpmn:outgoing>
      <bpmn:outgoing>Flow_0mnop78</bpmn:outgoing>
    </bpmn:exclusiveGateway>

    <bpmn:task id="Task_2" name="Validate Order">
      <bpmn:incoming>Flow_0ijkl56</bpmn:incoming>
      <bpmn:outgoing>Flow_0qrst90</bpmn:outgoing>
    </bpmn:task>

      ...

 </bpmndi:BPMNDiagram>
</bpmn:definitions>

```

---

# 🔧 Usage

```python
from datasets import load_dataset

ds = load_dataset("pritamdeka/BPMN-VLM")

example = ds["train"][0]

image = example["image"]               # PIL image object
bpmn_text = example["bpmn"]           # XML content as string
image_name = example["image_filename"]
bpmn_name = example["bpmn_filename"]
```

---

# 🎯 Applications

This dataset is suitable for:

- BPMN diagram understanding and parsing  
- OCR + VLM multimodal pipelines  
- Structured JSON extraction  
- Diagram-to-XML reconstruction  
- Fine-tuning Pixtral, Qwen2.5-VL, LLaMA 3.2 Vision, Aya Vision, Gemma3 and other VLMs 
- Evaluation against ground truth `.bpmn` files  

Ideal for research in:

- Vision-language reasoning  
- Diagram understanding  
- Business process modelling automation  

---

# 📜 Citation

If you use this dataset, please cite:

```bibtex
@misc{deka2025structuredextractionbusinessprocess,
      title={Structured Extraction from Business Process Diagrams Using Vision-Language Models}, 
      author={Pritam Deka and Barry Devereux},
      year={2025},
      eprint={2511.22448},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2511.22448}, 
}
```

---

# 📄 License

This dataset is released under **CC BY-NC 4.0** —  
It can be used for **research and non-commercial purposes** with attribution.

---

# 🙏 Acknowledgements

Developed at the **Advanced Research Centre (ARC), Queen’s University Belfast**, as part of research into multimodal structured extraction from business process diagrams.

---

# 📬 Contact

For questions, contact:  
**Pritam Deka** — *p.deka@qub.ac.uk*