There is no accuracy score here. The numbers are counts: characters (table markup included, so a table parser's count runs high) and the tokens it wrote, with how many are numbers. A dictionary-based score was removed because on this document it was measuring the wrong thing — page 4 is 77% numbers, and its remaining words are column abbreviations (“CLD”, “ABV”, “TDME”) that no dictionary contains, so a correct reading scored 30%. Judging accuracy needs a hand-transcribed ground truth, which these pages do not have. Compare each pane against the scan.