OmniDocBench

Run `npx skills add opendatalab/OmniDocBench --all -g -y` in your terminal to install the evaluation helper skill globally.

It is a benchmark for testing how well document parsing and OCR models handle real-world PDFs like papers, newspapers, and handwritten notes. It scores text, formulas, tables, layout, and reading order.

Point the skill at your ground-truth JSON and prediction markdown folder, and it generates a config and runs the evaluation for you. It then reports Overall, Text, Formula, and Table scores with result file paths.

Yes. Docker is the recommended way to run evaluations because it bundles the TeX, ImageMagick, and Ghostscript dependencies needed for CDM formula scoring.

Yes. The AI agent validates your file paths, builds the config, runs the evaluation, and explains the scores based on your plain-English requests.