Dany

Is it simply a matter of passing in a `writing_direction` here and flipping the text? https://github.com/koreader/koreader/blob/9d94ab38f34e60a3635607413e60868daef25d75/frontend/ui/widget/textboxwidget.lua#L502

Interestingly, the logs show the text in the correct direction. It isnt until it makes it into the display widget (and the dictionary query word) that it actually gets reversed...

If you can point me to some possible places where the text reversal might be happening that would be help me debug a bit more.

Thank you! > You could compare what is logged when selecting arabic in an EPUB (where I know everything should work) and in a PDF (where things are more obscure...

For what it's worth here is what I've done as a hack for myself locally: `frontend/document/pdfdocument.lua` ```lua function shouldreverse(str) return string.find(str, "[\216-\219][\128-\191]") ~= nil end function utf8reverse(str) local rev =...

Shouldnt this be easy to catch earlier up the call stack and emit a warning or error?

I'm running into this issue as well. I'm struggling to get helpful test errors from doo. I'd like to take a stab at this, @bensu do you have any pointers...

It looks like the code to train a new model has not been made available

Not a fix, but as a temporary solution you can probably point to my branch of catwalk which uses my branch of tango that allows newer torch versions: https://github.com/allenai/catwalk/tree/dh-debug According...

Ah, this actually isn't doing anything different from what the code was originally doing with the `TOLERANCE` constant. I'll reopen if I have a general fix.