Infinity-Doc-400K: 400,000 Scanned Documents with Structural Annotations

Infinity-Doc-400K contains 400,000 real-world and synthetic scanned documents featuring diverse layout variations and structural annotations, released by infly in 2025. This dataset expands upon the previous 55K version to support document parsing model training across financial, medical, and academic domains. It is provided in Parquet format and includes both image and text modalities.

Use Cases

• Document parsing using structural annotations to identify layout components like tables and headers
• OCR training on diverse document types including medical and financial reports
• Layout analysis for academic reports to segment text blocks from figures

Strengths

• 400,000 total records providing significant scale for document parsing
• Includes structural annotations for complex layout identification
• Covers specialized domains including medical and financial reports

Limitations

• Mix of synthetic and real-world data may introduce domain shift for models targeting only physical scans
• Specific column schema and label taxonomy are not explicitly detailed in the metadata

Provenance

Source

infly (Arxiv 250603197)

Collection Method

Combination of real-world document collection and synthetic generation

Freshness

Last updated November 2025.