Nine diacritics, two centuries of print.

First, the nine diacritics. Standard OCR engines trained on English corpora routinely confuse ą with a, ł with l, ó with o. A single missed ogonek can invert the meaning of a Polish sentence.

Second, the historical stock. Polish documents from 1791 to the early twentieth century frequently use gothic or fraktur typefaces. Modern OCR is trained on modern fonts; historical print requires either a dedicated model or a post-correction step.

Third, the PolEval 2021 task is deliberately not raw OCR — it is OCR post-correction. Given noisy OCR output, a Polish NLP model must recover the correct text. This rewards language understanding, not vision.

What CER does and does not say.

Character Error Rate is a tidy Levenshtein over a known reference. 2.1% means roughly two errors per 100 characters; it does not say where those errors are. On Polish documents, a stripped ogonek is a CER error but often a different word.

Our CodeSOTA Polish panel deliberately tests both ends. The Wikipedia and Real Corpus categories let a model lean on a Polish language prior — a dictionary, n-gram statistics, fine-tuning on Polish text. The Synth Random and Synth Words categories deny that prior; the model has only the image.

A ten-fold gap between the two is the signal we track. A model that reads characters, not the dictionary, is the one that will hold up on the next out-of-distribution document you hand it.