jpeglle/epstein-files-ocr-complete

Epstein Files — Complete OCR Dataset

This is a comprehensive, structured publication of the Epstein Files OCR dataset, significantly expanding upon the earlier Datasets 1-8 release.

Dataset Summary

This dataset contains page-level OCR output compiled from an extensive release of documents related to Jeffrey Epstein / the Epstein case.

Each row in this dataset represents one scanned PDF document from the original release using a proprietary automated OCR pipeline provided by Wild Ma-Gässli.

The dataset is designed for:

• Question answering
• Information retrieval
• Downstream NLP tasks such as named entity recognition (NER), entity linking, and relationship extraction.

Enhancements from Previous Versions

• Scale: This structured release covers 1,380,935 PDF documents, comprising over 2,700,000 total pages.
• Format: Restructured from individual .md files into a more efficient Parquet format.
• Document Linking: Each page retains its original document_id (e.g., EFTA00000001), resolving the limitation from earlier releases where pages could not be easily traced back to their source PDFs.

Supported Tasks

• Text retrieval / search (BM25, hybrid, dense retrieval)
• Question answering over retrieved context (RAG)
• Entity extraction (names, places, phone numbers, dates) from noisy OCR
• Social graph and network analysis

Languages

Primarily English (en).

Related Tools

This dataset is designed to be used with the Epstein Chat analysis tool, which provides a RAG (Retrieval-Augmented Generation) interface for querying these documents.

• GitHub Repository: ishumilin/epstein-chat

Dataset Structure

The dataset is provided as a Parquet file, which works natively with Hugging Face's datasets library.

Data Fields

The schema contains the following fields:

• document_id (string): The identifier of the original document/page (e.g., EFTA00146767).
• content (string): The full OCR-extracted content for that specific document.

Example Row:

{
  "document_id": "EFTA00146767",
  "content": "Hey beautiful. Tried to call you back..."
}

Splits

No predefined train/validation/test splits.

Dataset Creation

Source Data

• Primary source: The upstream Epstein Files release hosted at:
  • Torrent: https://github.com/yung-megafone/Epstein-Files/blob/main/Torrent%20Files/epstein-files-structured-full-20250204.tar.zst.torrent

Coverage in this dataset: All PDF files from the upstream release.

OCR / Preprocessing

OCR was performed on this dataset using a proprietary model provided by Wild Ma-Gässli.

Considerations for Using the Data

Personal / Sensitive Information

These documents contain personal data (names, phone numbers, addresses, emails) and/or information about alleged criminal activity.

Redaction Policy:

• This dataset is published as verbatim OCR output derived from the public source files.
• No additional redaction (masking/removal) has been applied beyond what was already redacted by the DOJ or the original releasing entity.

Use Responsibly:

• Comply with applicable laws and platform policies.
• Avoid doxxing or harassment.
• Do not treat OCR text as ground truth; always verify against the original page images/PDFs for high-stakes use.

Known Limitations

• OCR noise: While improved, automated extraction can produce recognition errors, incorrect formatting artifacts, or miss obscure characters (especially on poor-quality scans or handwriting). Some pages contain explicit placeholders such as [hidden text] reflecting original redactions made by DOJ.
• Content variance: Documents range from dense narrative text to unformatted tables and metadata tags.
• Corrupted Source Files: Three files from the original release were severely corrupted and their contents remain unknown and unextracted:
  • EFTA00645624.pdf
  • EFTA01175426.pdf
  • EFTA01220934.pdf

Biases

This dataset reflects:

• The selection, redaction, and presentation choices of the original releasing institution.
• OCR model performance characteristics (better on clean text, worse on handwriting / low-quality scans).

Licensing

See LICENSE for the full CC0 1.0 legal text.

Citation

If you use this dataset, please cite:

1. The original public release.
2. This dataset.