Lukaszl/pl-newspaper-pages-ocr-dataset-100-v1-results

---
license: mit
tags:
- ocr-bench
- leaderboard
source_datasets:
- Lukaszl/pl-newspaper-pages-ocr-dataset-100-v1
configs:
- config_name: default
data_files:
- split: train
path: data/train-*.parquet
- config_name: comparisons
data_files:
- split: train
path: comparisons/train-*.parquet
- config_name: leaderboard
data_files:
- split: train
path: leaderboard/train-*.parquet
- config_name: metadata
data_files:
- split: train
path: metadata/train-*.parquet
---
# OCR Bench Results: Polish newspaper and press-style pages benchmark

VLM-as-judge pairwise evaluation of OCR models on Polish newspaper, magazine, and press-style pages. Results depend strongly on document type, so this should be read as a document-specific benchmark rather than a universal OCR ranking.

This benchmark focuses on Polish press-style page images with visually complex layouts, including multi-column text, embedded images, side blocks, highlighted sections, and advertisement-like compositions. These pages are challenging for OCR not because of low image quality alone, but because they often require correct reading order reconstruction across non-trivial layouts. Even when individual words are recognized correctly, OCR quality can still break down if columns are merged incorrectly, side content is inserted in the wrong place, or text flow is reconstructed poorly.

## Leaderboard

| Rank | Model | Params | ELO | 95% CI | Wins | Losses | Ties | Win% |
|------|-------|--------|-----|--------|------|--------|------|------|
| 1 | clearocr.com/clearocr-api | | 1916 | 1865–1989 | 375 | 24 | 0 | 94% |
| 2 | lightonai/LightOnOCR-2-1B | 1B | 1529 | 1495–1566 | 221 | 178 | 0 | 55% |
| 3 | deepseek-ai/DeepSeek-OCR | 4B | 1504 | 1468–1542 | 208 | 190 | 0 | 52% |
| 4 | FireRedTeam/FireRed-OCR | 2.1B | 1387 | 1350–1420 | 147 | 252 | 0 | 37% |
| 5 | zai-org/GLM-OCR | 0.9B | 1163 | 1107–1208 | 46 | 353 | 0 | 12% |

## Details

- **Task**: OCR (Optical Character Recognition)
- **Language**: Polish
- **Document type**: Newspaper, magazine, and press-style page layouts
- **Original dataset**: [`Lukaszl/pl-newspaper-pages-ocr-dataset-100`](https://huggingface.co/datasets/Lukaszl/pl-newspaper-pages-ocr-dataset-100)
- **Source dataset**: [`Lukaszl/pl-newspaper-pages-ocr-dataset-100-v1`](https://huggingface.co/datasets/Lukaszl/pl-newspaper-pages-ocr-dataset-100-v1)
- **Judge**: Qwen3.5-35B-A3B
- **Comparisons**: 997
- **Method**: Bradley-Terry MLE with bootstrap 95% CIs

## About the dataset

The benchmark uses a small public image-only OCR dataset containing 100 Polish-language press / magazine-style page images sampled from a multilingual document collection.

The evaluated pages include layouts such as:

- article pages
- multi-column text layouts
- pages with embedded images
- pages with side blocks and highlighted sections
- advertisement-style pages
- editorial and magazine-like compositions

This is a useful stress-test for OCR systems that claim to handle real-world reading-order reconstruction rather than only plain single-column text.

## About clearOCR

[clearOCR](https://clearocr.com) is an OCR API for extracting text from PDFs, scans and document images, with a strong focus on **Polish and English documents**.

New accounts currently receive:

- **1,000 free single-image OCR runs**
- valid for **30 days**

API access is available via the clearOCR website:

https://clearocr.com

## Configs

- `load_dataset("Lukaszl/pl-newspaper-pages-ocr-dataset-100-v1-results")` — leaderboard table
- `load_dataset("Lukaszl/pl-newspaper-pages-ocr-dataset-100-v1-results", name="comparisons")` — full pairwise comparison log
- `load_dataset("Lukaszl/pl-newspaper-pages-ocr-dataset-100-v1-results", name="metadata")` — evaluation run history

*Generated by [ocr-bench](https://github.com/davanstrien/ocr-bench)*