OCR plus Jev throughput at a fixed review threshold

#1
by alphabell-research - opened

Congratulations on the release. The MiniCPM-OCR 1.0 card reports parser throughput up to 128 pages per minute. The Jev card reports AUROC and accuracy-versus-review-free curves, and describes scoring each parsed region. Those are useful separate measurements.

For the combined pipeline, could you publish a paired measurement on the same hardware and document set: parser-only pages/minute, OCR plus Jev pages/minute, mean regions per page, Jev batch size, threshold, false-accept rate and fraction sent to human review? A small table for receipts, legal pages and complex documents would connect your existing review-free curves to real capacity planning. Verified pages/minute at a stated error ceiling is the deployment number that neither model's standalone benchmark can supply.

Sign up or log in to comment