PDFA Eng Wds: 1,000+ English PDF Pages with Bounding Boxes

PDFA Eng Wds contains between 1,000 and 10,000 English document pages filtered from the SafeDocs CC-MAIN-2021-31-PDF-UNTRUNCATED corpus by pixparse. Released in 2024, it provides a curated subset specifically formatted for vision-language models with pre-processed bounding box annotations. The data serves as a machine-learning-ready version of the original SafeDocs document analysis collection.

Use Cases

• Training vision-language models to map visual regions to English text using bounding boxes
• Benchmarking document layout analysis tools on annotated page images
• Fine-tuning OCR models on real-world PDF layouts using the provided text labels

Strengths

• 1,000 to 10,000 records
• Includes bounding box annotations for document layout analysis
• Formatted for the WebDataset library to support efficient streaming

Limitations

• Small sample size of under 10,000 records
• Restricted to English language documents
• Temporal staleness as source data is from 2021

Provenance

Source

SafeDocs corpus (CC-MAIN-2021-31-PDF-UNTRUNCATED)

Collection Method

Filtered and processed from the Common Crawl PDF subset

Time Range

2021

Freshness

Last updated in March 2024; source material is derived from a 2021 Common Crawl.