Inconsistent results
#50
by TheMrguiller - opened
Hi,
I am testing GLM OCR and have gotten good results so far in my initial testing, but when moving to a real pipeline in Docling, I have gotten curious results. Docling layout models normally provide close-up crops of the tables, which affects the results of your models a lot, whereas adding padding improves the results. However, there are inconsistent results depending on how much padding I add (from 5 to 50), which should not be happening.
Is this something expected? Could i get information about the padding sizes of the images?
TheMrguiller changed discussion title from Incosistent results to Inconsistent results