Every PDF element, scored honestly.

OmniDocBench is the Shanghai AI Laboratory benchmark for end-to-end PDF document parsing. It grades text extraction, table structure, formula recognition and layout on the same page — which is why it decides who ships into an accounts-payable pipeline and who does not.

Composite = text + table + formula.

The headline OmniDocBench score is a composite defined by its authors as ((1 − TextEditDist) · 100 + TableTEDS + FormulaCDM) / 3. Each component tracks a different axis of parsing accuracy, so a model that aces tables but mangles formulas cannot win on the headline.

Secondary metrics tracked here: table-teds (tree edit distance on table recognition), formula-cdm (formula character-level match), layout-map (mean average precision on layout detection), and text-edit-distance / ocr-edit-distance (lower is better for edit distances). Every reported score is preserved verbatim from the submission.

A cross-section of real document wild-type.

OmniDocBench was released by Shanghai AI Laboratory as a comprehensive benchmark for evaluating PDF document parsing across diverse document types with multi-level annotations. It is the reference bench most open-source document parsers and vendor OCR APIs now report on.

The upstream leaderboard and test split live on alphaXiv. Per-row sources on the tables above link back to the submitting paper or vendor statement.

Reported, then reproduced.

Every row above is imported from the canonical benchmarks.json. Open-weight models are re-executed against the OmniDocBench test split through the CodeSOTA harness; closed APIs are run through the vendor endpoint with the model version and access date recorded. Rows marked “verified” have been independently reproduced, not taken from press release.

For the full reproduction policy, see the Codesota methodology.