---
dataset_info:
  features:
  - name: ID
    dtype: string
  - name: Height
    dtype: int64
  - name: Width
    dtype: int64
  - name: Entity_Masks
    dtype: string
  - name: Grouped_OCR_Blocks
    dtype: string
  - name: Caption
    dtype: string
  - name: Tags
    dtype: string
  - name: OCR_Text
    dtype: string
  - name: URL
    dtype: string
  - name: URL_sha256
    dtype: string
  - name: Language
    dtype: string
  - name: NSFW
    dtype: string
  splits:
  - name: train
    num_bytes: 1904454568
    num_examples: 120000
  download_size: 1144599398
  dataset_size: 1904454568
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# ADOPD: A Large-Scale Document Page Decomposition Dataset

[![ADOPD Tasks](.top_ADOPD_Tasks.png)](.top_ADOPD_Tasks.png)

**ADOPD** is a large-scale dataset designed for document image understanding. It introduces a novel data-driven document taxonomy discovery framework that combines large-scale pretrained models with a human-in-the-loop refinement process. The dataset supports four core tasks and includes rich annotations to foster progress in document analysis.

---

## 📊 Dataset Summary

- **Total Images**: 120,000  
- **Languages**:
  - English: 60,000  
  - Chinese: 20,000  
  - Japanese: 20,000  
  - Others: 20,000  

ADOPD ensures **diversity and balance** across document types and languages, validated through extensive experiments.

---

## 🧩 Supported Tasks

| Task       | Description                                | Field Name           |
|------------|--------------------------------------------|----------------------|
| **Doc2Mask** | Segment entity regions as pixel-level masks | `Entity_Masks`       |
| **Doc2Box**  | Detect and group OCR text blocks           | `Grouped_OCR_Blocks` |
| **Doc2Tag**  | Predict high-level semantic tags           | `Tags`               |
| **Doc2Seq**  | Generate abstracted captions               | `Caption`            |

---

## 📝 Annotation Fields

Each data sample includes the following fields:

- `ID` / `URL_sha256`: Unique identifier for each document image  
- `URL`: Direct link to download the image  
- `Height`, `Width`: Image resolution in pixels  
- `Entity_Masks`: Human-annotated segmentation masks for document entities  
- `Grouped_OCR_Blocks`: Grouped OCR text blocks with bounding boxes  
- `Caption`: Human-written descriptive caption summarizing the content  
- `Tags`: Predicted document-level semantic tags  
- `OCR_Text`: Raw plain-text extracted from the image  
- `Language`: Language of the document content  
- `NSFW`: Indicator flag for not-safe-for-work (NSFW) content  

---

## 📄 Citation

If you use ADOPD in your research, please cite:

```bibtex
@inproceedings{
    gu2024adopd,
    title={{ADOPD}: A Large-Scale Document Page Decomposition Dataset},
    author={Jiuxiang Gu and Xiangxi Shi and Jason Kuen and Lu Qi and Ruiyi Zhang and Anqi Liu and Ani Nenkova and Tong Sun},
    booktitle={The Twelfth International Conference on Learning Representations},
    year={2024},
    url={https://openreview.net/forum?id=x1ptaXpOYa}
}
```

---

## 📜 License

- **License**: [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/)  
- For **non-commercial use only**. Redistribution and derivative works are **not permitted**.