kenza-ily/publaynet_disco

PubLayNet_mini Dataset

A diverse mini subset of the PubLayNet dataset with 500 samples for document layout analysis evaluation.

Dataset Details

• Total Samples: 500 document images
• Source: PubLayNet training set (146,874 total samples)
• Task: Document Layout Analysis
• Format: Parquet with embedded images and annotations
• Image Size: 612×792 pixels (RGB)
• Categories: 5 layout element types

Categories

The dataset contains annotations for 5 categories of document layout elements:

1. Text (1): Regular text blocks and paragraphs
2. Title (2): Document titles and headings
3. List (3): Bulleted or numbered lists
4. Table (4): Tabular data structures
5. Figure (5): Images, charts, and diagrams

Features

Each sample contains:

• id: Unique document identifier
• image: Document image (PIL Image) - automatically loaded from embedded bytes
• annotations: List of layout element annotations with:
  • category_id: Element type (1-5)
  • bbox: Bounding box coordinates [x, y, width, height]
  • area: Area of the bounding box
  • iscrowd: Whether the annotation is for a crowd of objects
  • id: Unique annotation identifier
  • image_id: Reference to the document image
  • segmentation: Polygon segmentation mask

Data Storage

Images are stored as embedded bytes in the parquet file and automatically converted to PIL Images when loaded. This ensures:

• Self-contained dataset (no external image dependencies)
• Fast loading and processing
• Compatibility with HuggingFace datasets library

Category Distribution

This subset maintains diverse representation across categories:

• Text: ~3,676 annotations
• Title: ~1,000 annotations
• List: ~73 annotations
• Table: ~128 annotations
• Figure: ~172 annotations

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("kenza-ily/publaynet-mini")

# Each sample contains:
for sample in dataset['train']:
    print(f"Document ID: {sample['id']}")
    print(f"Number of layout elements: {len(sample['annotations'])}")
    
    # Access the image (automatically converted to PIL Image)
    image = sample['image']  # PIL Image object
    print(f"Image size: {image.size}")
    
    # Access annotations
    for ann in sample['annotations']:
        category = ann['category_id']
        bbox = ann['bbox']
        segmentation = ann['segmentation']
        print(f"Element {category}: bbox={bbox}")

Loading from Parquet

You can also load the data directly from the parquet file:

import pyarrow.parquet as pq
import pandas as pd
from PIL import Image as PILImage
import io

# Read parquet file
table = pq.read_table("publaynet_mini.parquet")
df = table.to_pandas()

# Convert images from bytes to PIL Images
def convert_image(img_data):
    if isinstance(img_data, dict) and 'bytes' in img_data:
        img_bytes = img_data['bytes']
        return PILImage.open(io.BytesIO(img_bytes))
    return img_data

df['image'] = df['image'].apply(convert_image)

# Access data
for idx, row in df.iterrows():
    image = row['image']  # PIL Image
    annotations = row['annotations']  # List of dicts

Citation

If you use this dataset, please cite both the original PubLayNet paper and the DISCO paper, which introduces this evaluation subset.

@inproceedings{zhong2019publaynet,
  title={PubLayNet: Largest Dataset Ever for Document Layout Analysis},
  author={Zhong, Xu and Tang, Jianbin and Yepes, Antonio Jimeno},
  booktitle={2019 International Conference on Document Analysis and Recognition (ICDAR)},
  pages={1015--1022},
  year={2019},
  organization={IEEE}
}

@inproceedings{benkirane2026disco,
  title={{DISCO}: Document Intelligence Suite for Comparative Evaluation},
  author={Benkirane, Kenza and Asenov, Martin and Goldwater, Daniel and Ghodsi, Aneiss},
  booktitle={ICLR 2026 Workshop on Multimodal Intelligence},
  year={2026},
  url={arxiv.org/abs/2603.23511}
}

License

This subset follows the original PubLayNet dataset license.