ishumilin/epstein-files-ocr-datasets-1-8-early-release

Epstein Files OCR — Datasets 1–8 (Early Release)

Work in Progress (WIP)

This is an early publication. We are actively working on improving OCR quality and expanding coverage.

Dataset Summary

This dataset contains page-level OCR output (as Markdown) from a public release of documents related to Jeffrey Epstein / the Epstein case.

Each Markdown file represents one scanned page converted to text using an automated OCR pipeline. The dataset is designed for:

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

High-level Statistics (This Repository Snapshot)

• File format: *.md
• Organization: Flat directory of pages
• Page files: page_<N>.md (e.g. page_10000.md)
• Number of page files: 42,182 (plus this README.md)
• Approx size (uncompressed): ~172 MB

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

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

Languages

Primarily English (en).

Dataset Structure

Folder Layout

To satisfy Hugging Face repository limits (max 10,000 files per directory), pages are stored in subfolders under pages/:

pages/
  00000-00999/
    page_1.md
    ...
  01000-01999/
  ...
  42000-42999/

Data Instances

Each file contains OCR-extracted text in Markdown. Content varies widely:

• Some pages are short identifiers (e.g. EFTA00000500).
• Some pages contain tables (e.g. call logs) rendered as Markdown tables.
• Some pages are narrative text.

Example (truncated):

DATE 7/6/19
CASE ID 31E-NY-3027571
FBI
...

Example (table excerpt, truncated):

| BILLED PHONE | CALL DATE | TIME | DIALED NUMBER | DEST NUMBER | ... |
| ... |

Data Fields

This dataset is distributed as raw text files and does not ship a structured Arrow/Parquet schema.

Typical fields you may want to derive during ingestion:

• page_number (integer parsed from filename)
• text (the Markdown content)
• source_id / document_id (not included in this flat export; see "Limitations")

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/combined-all-epstein-files_archive.torrent
  • PDF: COMBINED_ALL_EPSTEIN_FILES.pdf

Coverage in this dataset: Currently corresponds to datasets 1–8 from the upstream release.

OCR / Preprocessing

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

Output was saved as Markdown, one file per page.

Considerations for Using the Data

Personal / Sensitive Information

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

Redaction Policy:

• This dataset is published as verbatim OCR output derived from the upstream public PDFs.
• No additional redaction (masking/removal) has been applied beyond what is present in the upstream source materials.

Use Responsibly:

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

Some pages contain explicit placeholders such as [hidden text] reflecting original redactions made by DOJ.

Known Limitations

• OCR noise: Recognition errors, formatting artifacts, missing characters.
• No document grouping: In this snapshot, pages are stored as page_<N>.md without a stable link back to the originating PDF and page-in-PDF.
  • If you have access to the original PDFs, consider publishing a metadata.jsonl with {page_number, pdf_name, pdf_page_index, sha256, ...}.

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 (institution + date + release name)
2. The upstream archive (if applicable)
3. This OCR dataset (Hugging Face dataset URL once published)