davanstrien/europeana-photo-tasks-v1

Europeana Newspapers Sample Dataset

Dataset Description

This is a curated sample from the Europeana newspapers dataset, prepared for Vision-Language Model (VLM) experiments, particularly for testing structured metadata extraction from historical newspaper pages.

Dataset Sources

• Original Dataset: biglam/europeana_newspapers
• Source: Europeana digital library
• License: See original dataset

Dataset Structure

Each sample contains a newspaper page image downloaded via IIIF along with associated metadata.

Data Fields

• image: The newspaper page image (PIL Image)
• date: Publication date (ground truth for extraction tasks)
• title: Newspaper title
• issue: Issue number
• iiif_url: Original IIIF image URL
• text: OCR text content
• mean_ocr: OCR quality score (0-1)
• language: Document language
• issue_uri: Unique identifier
• download_status: Status of image download
• download_retries: Number of retry attempts needed

Use Cases

This dataset is ideal for:

• Testing VLM metadata extraction capabilities
• Training models with GRPO for structured information extraction
• Evaluating OCR and document understanding models
• Benchmarking on historical newspaper understanding

Example Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("your-username/europeana-newspapers-sample")

# Access a sample
sample = dataset['train'][0]
image = sample['image']  # PIL Image
date = sample['date']     # Ground truth date
title = sample['title']   # Ground truth title

Processing Notes

• Images were downloaded using IIIF protocol with automatic fallback to lower resolutions
• Failed downloads were excluded from the final dataset
• All images are converted to RGB format
• Dataset was created using async downloading for efficiency

Citation

If you use this dataset, please cite the original Europeana newspapers dataset:

@dataset{europeana_newspapers,
  title={Europeana Newspapers Dataset},
  author={BigLAM},
  year={2024},
  publisher={Hugging Face}
}