ferjorosa/bne-hemeroteca-ocr-xix

BNE Hemeroteca OCR Dataset (19th Century)

This dataset provides full-text OCR and high-resolution page images for 19th-century Spanish publications sourced from the Biblioteca Nacional de España (BNE) - Hemeroteca Digital. It consists of over 830,000 pages from approximately 40,000 documents, totaling roughly 800 million text tokens.

Note on Temporal Coverage: Although the selected publications originated in the 19th century, some long-running titles extend into the early 20th century. To apply a strict temporal filter, please use the date column.

Methodology

This dataset was built starting from the bne-hemeroteca-publications dataset. We filtered publications using the date column to focus on those from the 19th century, and selected 20 thematic collections (see full list below). Newspapers were intentionally excluded as their multi-column layouts did not perform well in manual OCR tests.

For each selected publication, we downloaded the corresponding issue PDFs from BNE, converted pages to images, and processed them with allenai/olmOCR-2-7B-1025-FP8 to extract text in markdown format.

Full list of collections included (20 collections)

Collection	Split
Agricultura y ganadería	agricultura_y_ganaderia
Ciencias	ciencias
Cultura	cultura
Deportes	deportes
Derecho	derecho
Economía	economia
Educación	educacion
Fuerzas armadas	fuerzas_armadas
Historia	historia
Industria	industria
Literatura	literatura
Medicina	medicina
Minería	mineria
Música	musica
Navegación	navegacion
Política	politica
Relaciones internacionales	relaciones_internacionales
Religión	religion
Revistas femeninas	revistas_femeninas
Teatro	teatro

Use Cases

• OCR benchmarking and fine-tuning: Use the existing OCR output as a baseline. Review and correct samples to generate ground truth for training improved OCR models on historical Spanish documents.

• Text retrieval and RAG: Since the original PDFs are scanned images, this dataset unlocks full-text search. Build retrieval-augmented generation (RAG) systems to query historical Spanish press.

• LLM pretraining: Use the generated text for general pretraining or to build time-capsule language models trained exclusively on 19th century text, capturing the language, topics, and worldview of that era.

Efficient Data Loading

If you are only interested in the text or metadata for early exploration (skipping the large images), we provide a script to download only specific columns efficiently.

• download_dataset.py

Fields

Page Data

Field	Description
image	Page image
text	Full OCR text (Markdown format)
issue_uuid	Unique identifier for the issue
issue_name	Name/Title of the issue
page	Page number
publication_issn	ISSN of the publication
date	Publication date
issue_link	Link to the original issue on BNE website
model_id	OCR model used (allenai/olmOCR-2-7B-1025-FP8)

OCR Metadata

Field	Description
primary_language	Primary language detected on page
is_table	Whether the page contains tables
is_diagram	Whether the page contains diagrams
is_rotation_valid	Whether page rotation was valid
rotation_correction	Degrees of rotation correction applied
total_input_tokens	Token count (input)
total_output_tokens	Token count (output)

Citation

@dataset{bne_hemeroteca_ocr_xix,
  title={BNE Hemeroteca OCR Dataset (19th Century)},
  author={Fernando Rodriguez},
  year={2025},
  url={https://huggingface.co/datasets/ferjorosa/bne-hemeroteca-ocr-xix},
  note={Processed with olmOCR from Hemeroteca Digital (BNE) archives}
}

License

CC0 1.0 (Public Domain). Original content provided by the Biblioteca Nacional de España.