Another test that directly concerns us in genealogy. Here is an example of an index generated by different administrative agents across different periods (over approximately a century).
The handwriting to decipher may be from the Prussian period (after 1872)?
Since I found this test too simplistic for our “machine intelligences,” I deliberately sought a document containing struck-through names and numbers (if there had been local place names in dialect, it would have been even better!).
In short, something that the human mind can decipher with a very low error rate, regardless of one’s education, culture, or language.
In the end, neither Transkribus with its French 1 model, nor Mistral OCR 3—which boasts extraordinary capabilities—managed to produce anything usable.
Basically, the idea was to provide a list of identifiers and associated names. I admit, I didn’t spend hours configuring these AIs. Just as I didn’t spend that time manually entering the data transcribed by my own brain…