I am currently running into the issue described in #712 (comment)
I guess the way I am creating the test data isn't quite correct (using createFromText and getDocumentPartsForLayoutTokens, the latter extracted from processShortNew).

Coverage increased (+0.03%) to 38.19% when pulling 6911ef2 on elifesciences:fix-fulltext-block-start into bfc10f7 on kermitt2:master.

I am currently running into the issue described in #712 (comment)
I guess the way I am creating the test data isn't quite correct (using createFromText and getDocumentPartsForLayoutTokens, the latter extracted from processShortNew).

I got around the problem by manually creating DocumentPiece for the whole document. I therefore reverted getDocumentPartsForLayoutTokens as it wouldn't be used by the test and could be refactored separately.

Other parsers are potentially affected as well.

There appear to be a lot of duplication and and could probably be refactored.

Raised suggestion for refactoring the features: #718

I'm having trouble understanding the whole picture here, so I'm not sure I can really review this part.

There is an example in the issue #712 - happy to add more information?

I have regenerated the fulltext model feature files with the fix, but actually there is no difference with before with respect to starting block. I also retrained the model (before checking the difference in features), in branch https://github.com/kermitt2/grobid/tree/fix-712
So apparently the problem in #712 never applies in practice in the training data at least.

It might appear in some new files like in your example, but the benchmark on PMC set only changes at the second decimal, so it should be minor.

So we could simply merge the fix, but no need to update the training files and model for this.