README.md

metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: label
      dtype: string
    - name: label_geometry
      dtype: string
  splits:
    - name: train
      num_bytes: 957239688
      num_examples: 1708
    - name: test
      num_bytes: 208067916
      num_examples: 410
  download_size: 1136749864
  dataset_size: 1165307604
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*

📄 BiblioPage

BiblioPage dataset — a collection of 2,118 scanned title pages annotated with structured bibliographic metadata. The dataset serves as a benchmark for bibliographic metadata extraction, document understanding, and visual language model evaluation.

🔍 About the Dataset

• 2,118 annotated title pages from Czech libraries (1485–21st century)
• 16 bibliographic attributes (e.g., Title, Author, Publication Date, etc.)
• Annotations include both text and precise bounding boxes
• Evaluation results available for YOLO, DETR, and VLLMs (GPT-4o, LLaMA 3)

You can download the dataset on Zenodo.
For more details, see the BiblioPage paper.