Olmocr Bench English: 6,483 Document OCR Test Cases

Olmocr Bench English is a document text extraction benchmark containing 6,483 English-only test cases across seven specialized categories including mathematical notation and complex layouts. Created by sarvamai as a filtered version of the allenai/olmOCR-bench dataset, it provides a targeted evaluation set for OCR models updated as of February 2026. The collection focuses on challenging document elements such as multi-column text, tables, and historical scans.

Use Cases

• Evaluating OCR accuracy on complex mathematical formulas using the arxiv_math and old_scans_math categories
• Benchmarking document layout analysis performance on multi_column and tables samples
• Testing text extraction reliability for small or dense typography using the long_tiny_text category

Strengths

• 6,483 total test cases across 7 distinct document categories
• High retention of 99.7% for mathematical documents from the original source
• Includes 526 old scans to test performance on degraded or historical media

Limitations

• Strictly English-only, making it unsuitable for multilingual OCR evaluation
• Significant reduction in the headers_footers category with 31.6% of original samples removed

Provenance

Source

allenai/olmOCR-bench

Collection Method

filtered from existing benchmark

Freshness

Updated in February 2026; reflects modern document processing challenges.