Dataset Open Access
This dataset is a CC-BY 4.0 subset of what bioRxiv kindly made available: https://www.biorxiv.org/tdm
It is randomized and split into train (6,000), validation (2,000) and test (2,000) subsets - 10,000 PDF / XML pairs in total.
The zip files further contain file lists of smaller subsets that used the subject area to potentially create a balanced subset.
The zip is similar in structure to the "PMC sample 1943" dataset that was created as part of: https://doi.org/10.1145/2494266.2494271 (a working link is available from: https://grobid.readthedocs.io/en/stable/End-to-end-evaluation/).
Constantin, A., Steve, P., Andrei, V.: Fully-automated PDF-to-XML conversion of scientific literature. In: Proceedings of the ACM Symposium on Document Engineering, pp. 177–180. ACM, New York (2013). doi: 10.1145/2494266.2494271