Olmocr Mix 1025: 270,000 PDF Pages with GPT-4.1 OCR Text

270,000 PDF pages converted to plain-text using GPT-4.1 with a specialized prompting strategy to preserve natural reading order and born-digital content. Released by the Allen Institute for AI (allenai) in October 2025, this collection includes both the original PDF files and their corresponding high-fidelity OCR outputs. It serves as an improved successor to the olmOCR-mix-0225 dataset with cleaner processing and more consistent formatting.

Use Cases

• Training OCR models to handle complex layouts using the paired PDF and text data
• Evaluating document parsing pipelines for natural reading order accuracy across 270,000 pages
• Fine-tuning multimodal models for document understanding using the page-to-text mappings

Strengths

• 270,000 PDF pages with paired high-quality text
• GPT-4.1 generated OCR outputs using specialized prompting
• Preservation of natural reading order and born-digital content

Limitations

• Potential for GPT-4.1 hallucinations or errors in text extraction
• Geographic bias toward US-based documents as indicated by metadata tags

Provenance

Source

Allen Institute for AI (allenai)

Collection Method

Synthetic/Annotated (GPT-4.1 OCR extraction from PDF sources)

Freshness

Last updated October 2025

Geography

United States