JamesConley/epstein-estate-documents-seventh-production

Epstein Estate Documents - Seventh Production

Dataset Description

This dataset contains the Epstein Estate Documents from the Seventh Production release with a separate OCR folder ran for all Images. OCR was completed using Deepseek OCR using vLLM.

Dataset Summary

• Total Documents: ~26,000+ files
• OCR Markdown Files: 23,124 documents
• Original Text Files: 2,897 documents
• Native Files: 11 files (Excel spreadsheets and video)
• Database Files: 2 files
• Source: House Oversight Committee - Epstein Estate Documents Production

Dataset Structure

epstein-estate-documents-seventh-production/
├── IMAGES/      # Original scanned images (22,903 .jpg, 221 .tif)
├── NATIVES/     # Native files (8 .xls, 2 .xlsx, 1 .mov)
├── TEXT/        # Original text documents (2,897 .txt)
└── OCR/         # OCR-processed markdown documents (23,124 .md)

The original dataset also includes a DATA folder containing binary files like

├── DATA/        # Database files (1 .dat, 1 .opt)

However the .dat was blocked from upload due to being a large binary file.

Data Fields

OCR Markdown Documents (OCR/)

• Format: Markdown (.md)
• Count: 23,124 files
• Organization: Organized in numbered subdirectories (001, 002, 006, 011, 012, etc.)
• Naming Convention: HOUSE_OVERSIGHT_XXXXXX.md
• Content: OCR-extracted text with markdown formatting, tables, and redaction markers

Original Text Files (TEXT/)

• Format: Plain text (.txt)
• Count: 2,897 files
• Content: Original text extractions, emails, correspondence

Native Files (NATIVES/)

• Formats: Excel (.xls, .xlsx), Video (.mov)
• Count: 11 files
• Content: Original spreadsheets and multimedia files

Original Images (IMAGES/)

• Formats: JPEG (.jpg), TIFF (.tif)
• Count: 23,124 images
• Content: Scanned documents, correspondence, financial records, flight logs, and other materials

OCR Methodology

Model Details

• Model: deepseek-ai/DeepSeek-OCR
• Framework: vLLM (for efficient batch inference)

OCR Configuration

Sampling Parameters:
- Temperature: 0.0 (deterministic output)
- Max Tokens: 8192
- N-gram Size: 30
- Window Size: 90
- Special Tokens: 
  - 128821: <td> (table cell open)
  - 128822: </td> (table cell close)

Software Stack Used:

vllm
Pillow
deepseek-ai/DeepSeek-OCR (auto-downloaded)

Source

Original documents released by the U.S. House Committee on Oversight and Accountability.

Contact

For questions or issues with this dataset, please open an issue in the dataset repository.

---

Note: Markdown files are purely generated by Deepseek OCR and have not been further validated. Users should verify critical information against original source documents when necessary.