Why the page header is often ignored in model's output?

[bryanyzhu]

Why the page header is often ignored in model's output?

[lukas-blecher]

yes, both actually:)

[lukas-blecher]

Thank you,

yes we removed the headers from the training data. And it works pretty well for pdfs however the scanned documents are out of distribution and might show unwanted artifacts.

That the fist line is missing is a bit weird. not 100% sure, but could be due to an unrelated issue with the page splitting algorithm.

[bryanyzhu]

Thanks, it indeed works really well. I'm curious to know the motivation of removing headers when preparing the training image-latex pair, is it beneficial for page splitting, or for easier post-processing, or something else? Appreciate it.