Key answer
Compare recognition against the original image, checking characters and relationships among rows, columns, footnotes and legends. Fluent extracted text does not establish faithful reconstruction.
Full explanation
OmniDocBench evaluates text, formulas, tables and reading order across nine document types, including papers, textbooks and slides. Its table experiments show effects from merged cells and rotation, with separate metrics for text recognition and table structure. Checking only whether any text was extracted is insufficient. Sections 4–5 and Table 7
Prioritize details that change judgments:
· Text and numbers: names, model numbers, negations, decimal points, signs, dates and units; flag unclear characters for confirmation.
· Tables: each value's row/column, multilevel headings and merged-cell scope; distinguish blank, zero and not applicable.
· Footnotes and continuations: which conclusion a note limits, whether continued tables connect correctly, and whether two-column text is interleaved.
· Charts: axes, scales, legends, colors, error bars and units. Distinguish values visually estimated from a plot from directly labeled data.
Transcribe the page 6 table first, retaining full headers, units and footnotes and flagging unclear cells. Do not calculate conclusions yet; compare groups after I confirm header/value alignment.
Provide clear originals and enlarged crops when needed, retaining relevant headers, axes and footnotes. Prefer original data tables for precise calculation. Matching totals and components are useful checks but can share the same copying error; inspect key values against the image.