sarvamai/olmOCR-Bench-English

olmOCR-bench (English Only)

This is a filtered version of the allenai/olmOCR-bench dataset containing only English documents.

Test Cases: Before vs After

Category	Before	After	Removed	% Retained
arxiv_math	2927	2917	10	99.7%
headers_footers	760	520	240	68.4%
long_tiny_text	442	442	0	100.0%
multi_column	884	691	193	78.2%
old_scans	526	526	0	100.0%
old_scans_math	458	458	0	100.0%
tables	1022	929	93	90.9%
TOTAL	7019	6483	536	92.4%

PDF Files: Before vs After

Category	Before	After	Removed	% Retained
arxiv_math	522	521	1	99.8%
headers_footers	266	190	76	71.4%
long_tiny_text	62	62	0	100.0%
multi_column	231	182	49	78.8%
old_scans	98	98	0	100.0%
old_scans_math	36	36	0	100.0%
tables	188	169	19	89.9%
TOTAL	1403	1258	145	89.7%

Summary

Metric	Original	English-Only	Retained
Test Cases	7019	6483	92.4%
PDF Files	1403	1258	89.7%

Usage

This dataset follows the same structure as the original olmOCR-bench:

├── README.md
├── stats.json
└── bench_data/
    ├── *.jsonl (test cases for each category)
    └── pdfs/<category>/*.pdf

Categories

• arxiv_math: Math formulas from arXiv papers
• headers_footers: Header/footer detection tests
• long_tiny_text: Long documents with tiny text
• multi_column: Multi-column layout documents
• old_scans: Historical scanned documents
• old_scans_math: Math from historical scans
• tables: Table extraction tests

License

This dataset inherits the ODC-BY-1.0 license from the original olmOCR-bench dataset.

Original Dataset

Based on allenai/olmOCR-bench.