Benchmark
OCRBench
Evaluate reading text in images, answering document questions, and extracting key information.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
The hosted original OCRBench subset contains 1,000 examples spanning text recognition, scene-text questions, document questions, key information extraction, and handwritten mathematics.
Methodology
Provide an image and its task question, then compare the answer against the accepted reference answers using the official evaluation rules. Inspect results by task type.