OmniParseBench
An open-source OCR document parsing evaluation benchmark released by Datalab. It uses a yes/no grading scheme rather than traditional scoring, and contains 16,288 test items covering 2,937 pages and 92 languages, supporting audits sliced by tag, writing system and language.
Timeline (last 90 days)
Datalab released the open-source OCR benchmark OmniParseBench
OmniParseBench contains 16,288 yes/no tests, covering 2,937 pages and 92 languages
OmniParseBench adopts a judging-based design, and the final score is a pass rate rather than a similarity score
Each score of OmniParseBench can be drilled down and sliced by tag, script, or language
OmniParseBench opens both the test items and the judging logic, so third parties can independently reproduce the same conclusions under the same inputs
The emergence of OmniParseBench pushes the competitive dimension of OCR evaluation from 'who scores higher' to 'who is reliable in specific scenarios'
All Facts (6)
OmniParseBench contains 16,288 yes/no tests, covering 2,937 pages and 92 languages
50%UnverifiedOmniParseBench adopts a judging-based design, and the final score is a pass rate rather than a similarity score
50%UnverifiedEach score of OmniParseBench can be drilled down and sliced by tag, script, or language
50%UnverifiedOmniParseBench opens both the test items and the judging logic, so third parties can independently reproduce the same conclusions under the same inputs
50%UnverifiedThe emergence of OmniParseBench pushes the competitive dimension of OCR evaluation from 'who scores higher' to 'who is reliable in specific scenarios'
50%UnverifiedDatalab released the open-source OCR benchmark OmniParseBench
50%