pupepps/epstein-emails

Epstein Email Threads Dataset

Dataset Summary

This dataset contains 5,082 parsed email threads extracted from OCR'd documents released by the U.S. House Oversight Committee. The emails have been processed using large language models to extract structured information including senders, recipients, timestamps, subjects, and message bodies, with OCR errors corrected and footers removed.

Dataset Description

Overview

This is a structured, machine-readable version of email threads extracted from the publicly released Epstein Estate documents. The original documents were OCR'd text files that required parsing to separate individual messages, extract metadata, and clean up OCR artifacts.

Key Features

• 5,082 email threads with structured metadata
• 16,447 individual messages across all threads
• Average of 3.24 messages per thread
• Chronologically ordered messages within threads
• OCR errors corrected
• Footers and disclaimers removed
• Quoted text properly separated from message bodies

Dataset Structure

Thread-level Schema (epstein_email_threads.parquet)

• thread_id: Unique identifier for the thread (e.g., "TEXT-001-HOUSE_OVERSIGHT_031683.txt_2")
• source_file: Original source filename from the House Oversight release
• subject: Thread subject line
• messages: JSON string containing list of messages in the thread
• message_count: Number of messages in the thread

Message Structure (within messages JSON)

Each message contains:

• sender: Sender name and/or email address
• recipients: Array of recipient strings
• timestamp: Message timestamp (various formats preserved)
• subject: Message subject line
• body: Message body content (quoted text and footers removed)

Source and Processing

Original Source

All documents originate from the public release "Oversight Committee Releases Additional Epstein Estate Documents" on the official House Oversight Committee website (press release dated November 12, 2025):

https://oversight.house.gov/release/oversight-committee-releases-additional-epstein-estate-documents/

The underlying materials are distributed via a Google Drive structure maintained by the Committee. This dataset is an independent derivative collection built from that release and is not an official product of the U.S. House of Representatives or the Committee on Oversight and Government Reform.

Source Dataset

This dataset is derived from the pre-processed CSV file available on Hugging Face:

https://huggingface.co/datasets/tensonaut/EPSTEIN_FILES_20K

The source dataset (tensonaut/EPSTEIN_FILES_20K) contains over 25,000 plain text files that were already OCR'd and organized into a single CSV file:

• TEXT/ – Files that were originally text-based (e.g., PDFs, emails) converted to plain text
• IMAGES/ – Image files (primarily JPG) converted to text via Tesseract OCR

The OCR processing was performed by the maintainers of the tensonaut/EPSTEIN_FILES_20K dataset. All image files (approximately 20,000 JPGs) were converted to machine-readable text using the open-source Tesseract OCR engine. Native text-based files were converted using standard PDF/text extraction tools.

Processing Pipeline

This dataset represents a further processing step on the tensonaut/EPSTEIN_FILES_20K CSV file:

1. Email Thread Parsing

Email threads were extracted from the pre-OCR'd CSV file using a large language model (LLM) with structured extraction:

• Model: xAI Grok 4.1 Fast via OpenRouter API
• Method: Structured extraction
• Process:
  • Identified email threads vs. non-email content
  • Separated individual messages within threads
  • Extracted headers (From, To, Sent, Subject)
  • Separated quoted replies from message bodies
  • Ordered messages chronologically (oldest first)
  • Removed footers, disclaimers, and signature blocks

2. Quality Improvements

Post-processing improvements were applied to enhance data quality:

• OCR Error Correction: Fixed common OCR mistakes (e.g., "jeeyacation" → "jeevacation")
• Footer Removal: Removed "HOUSE OVERSIGHT [number]" markers and legal disclaimers
• Quoted Text Cleanup: Removed quoted message citations from message bodies
• Recipient Extraction: Extracted missing recipients from message bodies where possible

3. Quality Assurance

• Automated Verification: All 5,082 threads validated for structure integrity
• Manual Review: Sample verification against original CSV source
• OCR Accuracy: Identified OCR errors corrected
• Data Integrity: No data loss or corruption detected

Usage and Responsibilities (Required Reading)

Intended Use

This dataset is provided for research and exploratory analysis purposes. Common use cases include:

• Academic research in fields such as computational linguistics, digital humanities, and social sciences
• Text analysis and natural language processing research on real-world email communications
• Exploratory data analysis of publicly released documents
• Qualitative research by journalists, historians, legal scholars, and researchers
• Methodological development for processing and analyzing structured email data

User Responsibilities

Users are responsible for:

• Using the dataset only for lawful purposes and in accordance with institutional and ethical review requirements
• Treating individuals mentioned in the documents with respect, and avoiding sensationalism or misuse of sensitive material
• Clearly distinguishing model-generated content and exploratory findings from verified facts, and citing primary sources where appropriate
• Complying with applicable law, institutional policies, and the terms of the original House releases

Not Intended For

This dataset is not intended for:

• Fine-tuning language models
• Harassment, doxing, or targeted attacks on any individual or group
• Attempts to deanonymize redacted information or circumvent existing redactions
• Making or amplifying unverified allegations as factual claims

All use must comply with applicable law, institutional policies, and the terms of the original House releases. See the "Legal and Copyright Status" and "Ethical and Content Warning" sections below before working with this corpus.

Legal and Copyright Status (Non-Authoritative)

The original underlying documents were created by various private individuals and entities, not by the dataset maintainer.

The documents are sourced from releases published by the U.S. House Committee on Oversight and Government Reform. The release webpages themselves carry standard copyright notices (© 2025 Committee on Oversight and Government Reform), and many individual documents are likely protected by copyright held by their original authors or rights holders.

This dataset:

• Does not assert any ownership over the underlying documents
• Does not grant any license to reproduce, distribute, or create derivative works from the underlying texts beyond what may already be permitted by law (e.g., fair use or similar doctrines in your jurisdiction)

Users are solely responsible for ensuring that their use of this corpus complies with applicable copyright law, privacy law, institutional policies, and the terms of the original House releases.

Nothing in this dataset card constitutes legal advice. If you plan to use this corpus in a public-facing product, for model training, or at scale, you should seek independent legal counsel.

Ethical and Content Warning

The documents contain material related to:

• Sexual abuse and exploitation
• Trafficking
• Violence and other highly sensitive topics
• Unverified allegations, opinions, or speculation

Users should be aware of the sensitive nature of this content and handle it appropriately in their research and analysis.

Data Quality

Extraction Accuracy

Based on sample verification against the original CSV source, the email extraction process was accurate:

• No hallucinations detected: All extracted content matches the source material
• No data loss: All messages and metadata were correctly extracted
• Accurate parsing: Message boundaries, headers, and bodies were correctly identified and separated

Note: While sample verification showed no hallucinations or data loss, errors may exist in the full dataset. If you encounter any extraction errors, hallucinations, or data inconsistencies, please report them by opening an issue on the Hugging Face dataset repository.

Known Limitations

• Source OCR Quality: Some original OCR errors from the source material may remain if not clearly identifiable (these are inherited from the tensonaut/EPSTEIN_FILES_20K dataset, not introduced during extraction)
• Timestamp Formats: Various timestamp formats preserved (not normalized) to maintain original format
• Email Addresses: Some email addresses may be incomplete or malformed due to source OCR, but were extracted as-is from the source material

Dataset Statistics

• Total Threads: 5,082
• Total Messages: 16,447
• Average Messages per Thread: 3.24

Citation

If you use this dataset in your research, please cite:

@dataset{epstein_emails_2025,
  title={Epstein Email Threads Dataset},
  author={notesbymuneeb},
  year={2025},
  url={https://huggingface.co/datasets/notesbymuneeb/epstein-emails},
  note={Derived from tensonaut/EPSTEIN_FILES_20K, which contains OCR'd text from U.S. House Oversight Committee public release}
}

Acknowledgments

This dataset is derived from:

1. Original Source: Publicly released materials by the U.S. House Committee on Oversight and Government Reform

   • https://oversight.house.gov/release/oversight-committee-releases-additional-epstein-estate-documents/

2. Intermediate Source: The pre-processed CSV dataset on Hugging Face

   • https://huggingface.co/datasets/tensonaut/EPSTEIN_FILES_20K
   • This dataset provided the OCR'd text files organized in CSV format

This dataset represents a structured, machine-readable version of email threads extracted from the tensonaut/EPSTEIN_FILES_20K CSV file.

Contact

For questions, issues, or feedback about this dataset, please open an issue on the Hugging Face dataset repository:

https://huggingface.co/datasets/notesbymuneeb/epstein-emails

---

Remember: This dataset contains sensitive content. Use responsibly and in accordance with all applicable laws and ethical guidelines.