The benchmarks provides TeX source files, PDF files and ground truth files for in total 12,099 scientific articles of arXiv.org. The files are given in folders src, pdf and groundtruth.
In the dataset of arXiv, scientific articles are organized in folders YYMM, where YY is a two-place digit representing the year and MM is a two-place digit representing the month of publication. For example, folder 1507 contains all articles from July 2015.
The folders src, pdf and groundtruth reflect this structure.
For illustration, consider the article astro-ph0001196 from January 2001. The associated files/folders are:
pdf/0001/astro-ph0001196.pdf, the PDF file.src/0001/astro-ph0001196, the folder with TeX source file and supplementary files like images, etc.groundtruth/0001/astro-ph0001196.body.txt, the ground truth file.
All of the 12,099 scientific articles were selected randomly, 1% from each YYMM folder, in order to represent the variety of topics and creation times (and thus article formats), which can be deduced from the arXiv submission rate statistics.
All ground truth files were generated using the benchmark-generator.
Each is a plain text file and contains the title, the section headings and the body text paragraphs of the associated scientific article, separated by blank lines and in order as they appear in the TeX file.
Each ground truth files has file extension .body.txt.
Please note that the provided PDF files are not those provided by arXiv, due to occasional (contentual) mismatches with the corresponding TeX files, but are regenerated from the provided TeX files.