CzechCore
Offline Czech grammar correction
- area
- AI and language
- stack
- Python · PyTorch · CTranslate2 · UDPipe
- as of
- Sep 2026
Problem
Correct Czech grammar on the device, with exact edit ranges an editor can underline, without sending text anywhere.
Approach
A 60.8M-parameter Czech correction model from ÚFAL reads the whole sentence, UDPipe provides exact token ranges, and deterministic guards protect URLs, e-mail addresses, dates, version and phone numbers, and code-like tokens. A separate local-LLM rewrite mode always asks for human review. The source includes an Android keyboard and a Windows TSF adapter.
Muži přišly domů.
Muži přišli domů.
Result
F0.5 of 0.634 (precision 0.788, recall 0.355) on the GECCC development split of 8,478 sentences, at about 229 ms per sentence on CPU. Released as a research alpha, because it still changes 558 of 3,288 correct sentences, too many for silent autocorrect.