README.md

metadata

license: cc-by-4.0
language:
  - sq
tags:
  - newspaper
  - historical
  - kosovo
  - albanian
  - ocr
  - vision-language
  - document-understanding
  - digital-humanities
pretty_name: Bujku Newspaper — Vision-Language OCR (1991–1998)
size_categories:
  - 100K<n<1M
task_categories:
  - image-to-text
  - image-text-to-text
modality:
  - image
  - text
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*

Bujku Newspaper — Vision-Language OCR Dataset (1991–1998)

An article-level image–text dataset for training and evaluating OCR and vision-language models on the Kosovo Albanian newspaper Bujku. Each row pairs an image crop of a single article from the newspaper with its OCR text and the text's token length.

• Rows: 113,300 segmented articles
• Years covered: 1991–1998
• Language: Albanian (sq)
• Modalities: image + text

About the newspaper

Bujku originated as a publication focused on agricultural topics and rural life. During the 1990s, following changes to Kosovo's institutional and media landscape that left few Albanian-language outlets in operation, Bujku expanded into a daily newspaper covering the broader news of the period. For much of the decade it was one of the main Albanian-language daily newspapers in Kosovo. Many of its journalists had previously worked at Rilindja, a large Kosovo media outlet that had ceased publication.

The newspaper is a valuable primary source for the history, language, and journalism of Kosovo in the 1990s. The scanned issues were digitized and published by the NGO FLOSSK (Prishtina, Kosovo) at https://books.flossk.org/gazetat/.

Dataset structure

Each record is a single segmented newspaper article.

Column	Type	Description
image	Image	An image crop of a single article from the newspaper, stored as an embedded image.
text	string	The OCR text of the article.
text_token_length	int64	Number of tokens in text (useful for filtering, batching, and length analysis).

Loading

from datasets import load_dataset

ds = load_dataset("Kushtrim/bujku_vl_ocr", split="train")

ex = ds[0]
ex["image"]              # PIL.Image of the page
print(ex["text"][:500])  # transcribed text
print(ex["text_token_length"])

Note: this dataset is gated — request access on the dataset page and authenticate (huggingface-cli login) before loading.

How the dataset was created

1. Data collection — Scanned PDF issues were obtained from the FLOSSK platform (https://books.flossk.org/gazetat/), organized by date as YYYY/BU-YYYYMMDD/.
2. Segmentation — Pages were segmented into individual articles, and each article was saved as its own image crop.
3. OCR — Each article image was processed with OCR to produce a clean text version.
4. Token length — text_token_length was computed over the transcribed text for each article.
5. Packaging — Images and text were assembled into a HuggingFace datasets.Dataset and exported as optimized Parquet.

Intended uses

• Training and benchmarking OCR and vision-language / image-to-text models on historical, low-resource (Albanian) print material.
• Document understanding and layout/transcription research on degraded historical scans.
• Digital humanities and cultural-heritage preservation.

For a text-only version of this archive with extracted entities, topics, and semantic embeddings (no images), see the related dataset Kushtrim/bujku.

Limitations

• Historical scan quality: source scans are of variable quality; transcriptions may contain errors, especially on degraded, faded, or low-contrast articles.
• Language & orthography: 1990s Albanian print may include older spellings, abbreviations, and diacritics that affect transcription fidelity.

Source and licensing

• Original scans: FLOSSK — https://books.flossk.org/gazetat/

Acknowledgements

Thanks to FLOSSK for scanning and openly publishing the Bujku archive, preserving an important part of Kosovo's cultural and historical heritage during one of the most sensitive periods in its history.