I have often wondered about this: does padding and chopping to context size along arbitrary boundaries have lasting effect on the overall data (due to misalignm

I have often wondered about this: does padding and chopping to context size along arbitrary boundaries have lasting effect on the overall data (due to misalignment at the ends, some words or lines getting chopped in the middle, etc.) Does that net out, because we do it consistently, or do we lose vital information that's compounded across the entire training data set?