Lukaszl/pl-government-docs-mix-ocr-dataset

Polish municipal administrative documents OCR dataset

An OCR-oriented image dataset built from publicly available Polish municipal administrative documents.

The dataset contains pages collected from materials such as:

• resolutions (uchwały)
• ordinances / regulations
• annexes
• official notices
• tabular administrative pages
• stamped and signed office documents
• other municipal administrative materials

All files in this repository are provided as images only.

What is inside

This dataset focuses on real-world Polish public-sector documents that often include:

• official text layouts
• stamps and seals
• handwritten notes
• signatures
• scan artifacts
• occasional tables
• mixed administrative formatting

The pages come from administrative documents published by 3 Polish municipalities.

Processing

Source files included formats such as:

• PDF
• JPG / JPEG
• GIF

They were normalized into a consistent image-only OCR evaluation format.

Processing steps:

• empty pages removed
• files converted to .jpg
• longer side resized to max 2100 px
• JPEG quality set to 95
• EXIF metadata removed

Intended use

This dataset is intended for:

• OCR benchmarking
• OCR regression testing
• Polish-language OCR evaluation
• comparison of OCR APIs and VLM-based OCR systems
• testing OCR robustness on real public administrative documents

It is especially useful for checking how OCR handles:

• Polish diacritics
• stamps and signatures
• noisy scans
• official layouts
• mixed document structure

Limitations

This is a document-specific OCR dataset.

Results obtained on this set should be interpreted in the context of:

• Polish municipal administrative documents
• image-only OCR evaluation
• a dataset collected from 3 municipalities

This is not a universal benchmark for all document types.

Source note

The dataset was built from publicly accessible Polish municipal administrative documents published online by local institutions.

This repository distributes processed page images for OCR evaluation and research purposes.

Maintainer

clearOCR https://clearocr.com