a hangry mouse @oscillatory.net · Apr 8
1 / 2

I finetuned Gemma 4 E2B on Math OCR to Typst. (the largest model I was able to run on my meager setup always outputs LaTeX). not bad, but it's chopping off the third line. first two lines are correct though. guessing it's due to single equation training data i used.

3 likes 1 replies

?

Replies

a hangry mouse · Apr 8

next up: better training data (multi-line, text + math, more canonicalization in labels), and maybe finetune from deepseek-ocr 2 instead?