OCR Arena

About OCR Arena

OCR Arena is a blind head-to-head evaluation playground for optical character recognition models, built by the team at Extend. You upload a PDF, JPEG, or PNG — or pull a random document from their sample pool if you do not have one handy — and two anonymous models extract the text side by side. You read both outputs, vote for the better one, and only then are the model identities revealed. Those votes feed an ELO leaderboard with win rate and battle counts per model, which is a far more honest signal for document work than vendor-published accuracy figures, because the documents being tested are whatever real users happen to throw at it: scanned invoices, handwriting, multi-column layouts, tables, low-contrast faxes, and non-Latin scripts. Beyond the battle mode there is a playground for running a single document against a chosen model when you already know what you want to test and just need to see the extraction quality, and an About page describing the methodology. For anyone choosing between a general vision-language model and a purpose-built OCR engine for a document pipeline, it is a fast way to sanity-check assumptions on your own representative files before committing to an API contract.

Key Features

Tags