I tested them by sending the 6 documents, each meant to represent a different document type, through my own webapp and comparing every output against the source. All ran on the same L4 GPU. The documents: Financial statements with merged multi-level headers (A typical annual report) Two pages of a two-column arXiv paper ("Deep Residual Learning for Image Recognition") Scanned German invoice with no text layer French municipal report with an embedded bar chart Typical data...