Ukraine · 2026-11-12

Ukraine Launches First Open Leaderboard for Ukrainian Handwriting Recognition

Ukrainian experts created a public table where anyone can compare which AI models read handwritten Ukrainian documents best, from archives to school notebooks.

✨ What happened

RUKOPYS Leaderboard, the first open leaderboard for models that recognize handwritten Ukrainian text, has appeared on Hugging Face. Such programs turn handwriting into regular printed text that can be searched and copied.

The initiative grew out of the spring Handwritten to Data competition on Kaggle. Its goal is to create a transparent standard for testing such systems.

📌 What's inside

The test is based on RUKOPYS, a large open dataset of handwritten Ukrainian texts created together with the Ukrainian Catholic University (UCU). The test uses 385 pages with 8,290 text regions. The documents span 1919 to 2025: archival records, sheets from the national unity radio dictation, university exams and school notebooks.

So far 43 models and systems have been tested in 94 configurations. They include commercial services accessed via API, open-weight models and specialized systems from competition participants. Evaluation covers accuracy, character error rate and the approximate cost of processing 1,000 pages.

🔹 Who came first

Gemini 4 Argon from Google ranks first on the composite score with 0.9387. Processing costs about $53.08 per 1,000 pages. Other Gemini versions follow, along with solutions from Anthropic and OpenAI.

The organizers say high-quality recognition does not always require high spending. Some commercial models on the cheapest settings almost match the most expensive ones, and compact specialized systems from participants compete with large general-purpose models.

🔸 Why it matters

The best models are already being used to digitize Ukrainian state archives, so old documents can become electronic and easier to find and read. The team is also working on a new generation of models that will combine the best participant solutions with a much larger training dataset.

All materials are publicly available for independent review: the model responses for each page, evaluation code, prompts and exact run settings. The developers invite the AI community to add their own models to the comparison.

Sources: 🔗 DOU/dev.ua