A frontier multilingual benchmark for table extraction from document images.
PulseBench-Tab contains 1,820 human-annotated tables across 9 languages and 4 scripts (Latin, CJK, Arabic, Cyrillic), sourced from 380 unique documents including financial filings, government reports, corporate disclosures, and regulatory filings. Each sample is a table image paired with a ground truth HTML table that preserves full structural markup (rowspan, colspan, thead, tbody).
