I work with a library (Biblioteca Philosophica Hermetica) in Amsterdam that has thousands of manuscripts from the renaissance to the early modern period… all very esoteric. We really want to get the renaissance into model training! Over 75% of books (1450-1700) are unscanned — and the manuscripts are in even worse shape.

Curious if anyone knows if there any new handwriting recognition benchmarks? I’ve noticed the main model providers have plateaued in the past year on their ability to read manuscripts / modern handwriting… I think the lack of well-designed competitive benchmarks is the issue…

I love positive examples of the intersection of AI and the humanities.

  As office boy I made such a mark
  That they gave me the post of a junior clerk.
  I served the writs with a smile so bland,
  And I copied all the letters in a big round hand—
      I copied all the letters in a hand so free,
      That now I am the Ruler of the Queen’s Navee!

https://robert-pfeffer.net/schriftarten/englisch/index.html?...

It was not very legible. Its legibility is weak due to aspects such as:

* Non decomposability into rectangles bounding letters

* Variation in letter forms depending on preceding and succeeding letters

* "extended serifs" or edges of letters which may vary at the author's pleasure.

* Some pen strokes being extremely thin to the point of near-invisibility (like the middle line of 'e' characters)

* non-uniform vertical height of letters (even ignoring ascenders)

* non-uniform horizontal baseline

1. When I loaded the page, it bombarded me with a banner asking me, "Interested in sports?" (Yes, I am, but I came here to read about English handwriting. Go away.)

2. At the end, it presented me with a "badge" for finishing a whole "book"! Yeah, maybe people's attention spans would be better if they weren't bombarded with little banners at the beginning.

/S