Hugging Face's logo

Datasets:
to-be
/
epstein-emails

Tasks:
Text Classification
Text Generation
Question Answering
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
1K - 10K
Tags:
text
email
legal
sqlite
vision-llm
structured-data
Libraries:
Datasets
pandas
License:
Dataset card Data Studio Files
xet
Community
1
epstein-emails / README.md
to-be's picture
to-be
Upload README.md
4214117 verified
preview code
|
raw
history blame contribute delete
11.5 kB
metadata
language:
  - en
pretty_name: Epstein Phone Messages Dataset
tags:
  - text
  - email
  - legal
  - sqlite
  - vision-llm
  - structured-data
  - datasets
license: other
task_categories:
  - text-classification
  - text-generation
  - question-answering
size_categories:
  - 1K<n<10K

Epstein Email Messages Dataset

Dataset Summary

This dataset contains 4,272 individual email messages extracted directly from screenshot images using advanced Vision LLM technology. Unlike other datasets that work with only pre-OCR'd text, this dataset also used the original JPG screenshot images from the U.S. House Oversight Committee release using Qwen 2.5 VL 72B vision model to extract structured email data with high accuracy.

The dataset powers a live Progressive Web App at epsteinsphone.org that provides an authentic Smartphone Messages-style interface for easy exploration of the emails, making it handy for researchers and journalists.

🚀 Live Demo & Exploration

🌐 Explore the Dataset Live at epsteinsphone.org

  • Smartphone Messages Interface: Authentic messenger-style conversation view
  • Full-Text Search: Search across all messages, subjects, and contacts
  • Offline Capable: Progressive Web App that works without internet
  • Mobile Optimized: Designed for phone-first exploration
  • Contact Profiles: Profile pictures and chronological conversation threading
  • Source Linking: Direct links to original screenshot documents

Dataset Description

Overview

This dataset was created for a complete end-to-end solution from raw government-released screenshot images to a fully functional web application. The processing pipeline uses Vision LLM technology to extract structured email data directly from image files, bypassing traditional OCR limitations and achieving higher accuracy in email parsing.

Key Technicalities

  • Vision-First Processing: Direct image-to-structured-data extraction using Qwen 2.5 VL 72B
  • Production Application: Live web app with 3,997 optimized messages (after deduplication and cleanup)
  • Open Source Pipeline: Complete processing methodology available on GitHub
  • Client-Side Database: SQLite database optimized for WebAssembly browser execution
  • Progressive Web App: Installable, offline-capable application with service worker

Dataset Structure

  • Parquet File (emails.parquet) - 1.1MB, flattened format optimized for data analysis and ML workflows

Database Schema

Two-table relational design preserving full email thread context:

-- Email document metadata table
CREATE TABLE email_documents (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    filename TEXT NOT NULL UNIQUE,
    subject TEXT,
    num_pages INTEGER,
    file_size_kb REAL,
    processed_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
    processing_status TEXT DEFAULT 'pending',
    error_message TEXT
);

-- Individual messages within email documents  
CREATE TABLE messages (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    email_document_id INTEGER NOT NULL,
    source_filename TEXT,
    subject TEXT,
    message_order INTEGER NOT NULL,
    from_address TEXT,
    to_address TEXT,
    other_recipients TEXT,          -- JSON array of CC/BCC recipients
    timestamp_raw TEXT,             -- Original timestamp format
    timestamp_iso TEXT,             -- Normalized YYYYMMDDHHMMSS
    message_html TEXT,
    document_id TEXT,
    FOREIGN KEY (email_document_id) REFERENCES email_documents(id)
);

-- Optimized indexes
CREATE INDEX idx_messages_timestamp ON messages(timestamp_iso);
CREATE INDEX idx_messages_filename ON messages(source_filename);
CREATE INDEX idx_messages_subject ON messages(subject);

Message Structure

Each message contains:

  • email_document_id: Foreign key linking to email_documents table
  • source_filename: Original source file from House Oversight release
  • message_order: Sequence number within the email thread
  • from_address: Sender name/email
  • to_address: Primary recipient
  • other_recipients: CC/BCC recipients as JSON array (e.g., ["Darren Indyke", "Debbie Fein"])
  • timestamp_raw: Original timestamp format preserved (e.g., "5/7/2019 4:29:00 AM")
  • timestamp_iso: Normalized ISO format (YYYYMMDDHHMMSS)
  • subject: Email subject line
  • message_html: HTML-formatted message body with preserved links
  • document_id: Source document identifier (HOUSE_OVERSIGHT_XXXXXX)

Key Differentiators

vs. notesbymuneeb/epstein-emails

Feature This Dataset notesbymuneeb/epstein-emails
Source Processing Vision LLM from screenshot images and pre-OCR'd emails Text LLM from pre-OCR'd emails
Data Structure Individual messages (flat structure) Threaded conversations (nested JSON)
Message Count 4,272 messages (filtered, deduplicated & optimized) 16,447 messages (comprehensive)
Processing Model Qwen 2.5 VL 72B (vision + text) xAI Grok 4.1 (text only)
Reproducibility Open-source pipeline on GitHub Processing methodology documented

Source and Processing

Original Source

All data originates from the official U.S. House Committee on Oversight and Government Reform release:

"Oversight Committee Releases Additional Epstein Estate Documents" (November 12, 2025)

Processing Pipeline

1. Vision LLM-Based Extraction

Model: Qwen 2.5 VL 72B via custom endpoint Method: Direct image processing with structured output Process:

  • Email documents identified from TEXT files starting with "From:"
  • Images resized to 800px height for optimal processing
  • Images were supplemented by the pre-OCR'd TEXT files
  • Base64 encoded for API submission
  • Vision model extracts: sender, recipient, timestamp, subject, HTML content
  • Parallel processing with 20-worker ThreadPoolExecutor

2. Data Quality Enhancement

  • Name Standardization: "Larry Summers" → "Lawrence Summers"
  • Jeffrey Epstein Variants: "jeeyacation", "jeevacation", "J" → "Jeffrey Epstein"
  • Timestamp Normalization: Multiple formats converted to ISO YYYYMMDDHHMMSS
  • Duplicate Detection: 90%+ similarity detection within 24 hour window (to handle different timezones)

3. Application Optimization

  • Database Optimization: Indexed for efficient web application queries
  • Message Cleanup: 4,272 final messages
  • Profile Integration: Conversation threading
  • Source Linking: Document IDs link to original JPG files

Dataset Statistics

  • Total Messages: 4,272
  • Email Documents: 1,951 (multi-message email threads)
  • Unique Senders: 676
  • Unique Recipients: 440
  • Unique Subjects: 795
  • Date Range: July 10, 2008 - July 6, 2019
  • Messages with CC/BCC: Contains JSON-structured additional recipients
  • Average Thread Length: ~2.2 messages per email document
  • Key Contacts: Jeffrey Epstein, Lawrence Summers, Lisa New, Steve Bannon, Ken Starr, Robert Kuhn, Ehud Barak, Boris Nikolic, Michael Wolff, Kathy Ruemmler and others

Usage and Applications

Intended Use Cases

  • Interactive Exploration: Use the live web app at epsteinsphone.org for intuitive message browsing
  • Academic Research: Computational linguistics, digital humanities, legal studies
  • Technical Research: Vision LLM evaluation, email parsing methodologies
  • Application Development: Fork the open-source codebase for similar projects
  • Data Analysis: SQL queries on structured email data
  • Timeline Analysis: Chronological relationship mapping

Open Source & Reproducibility

Complete processing pipeline available on GitHub: github.com/Toon-nooT/epsteins-phone-reconstructed

The repository includes:

  • data_processing.ipynb: Complete vision LLM processing pipeline
  • Web application source: Full PWA implementation
  • Configuration files: API setup and processing parameters
  • Documentation: Detailed methodology and setup instructions

Data Quality & Limitations

Strengths

  • High Accuracy: Vision LLM processing preserves visual context and layout information
  • Source Fidelity: Direct image processing achieves higher original document integrity
  • Comprehensive Metadata: Rich extraction including HTML formatting and source links

Known Limitations

  • Processing scope: Currently limited to emails ≤5 pages
  • Redacted Data: ~2,517 messages contain unresolved <REDACTED> fields
  • Duplicate Edge Cases: ~2,465 duplicate messages from complex threading
  • Timestamp Variations: Some inconsistencies from timezone differences
  • Incorrect Source Page: Sometimes document_id may not point to the exact page of the email thread

Legal, Ethical & Content Considerations

Legal Status

Source: Official U.S. House Committee on Oversight and Government Reform public releases Purpose: Research, educational, and informational use only No Affiliation: Independent project, not endorsed by any government entity, newspaper, media organization, academic institution, or other organization

Content Warning

This dataset contains sensitive material including:

  • References to legal proceedings and investigations
  • Personal communications and private correspondence
  • Names and contact information of public and private individuals
  • Discussions of business, political, and personal matters

User Responsibilities

Users must:

  • Use only for lawful purposes in accordance with applicable laws and institutional policies
  • Respect privacy of individuals mentioned in correspondence
  • Distinguish analysis from fact and cite primary sources appropriately
  • Avoid harassment or targeting of any individuals
  • Comply with terms of original House committee releases

Not Intended For

  • Model fine-tuning on personal communications
  • Harassment, doxing, or targeting individuals
  • Circumventing redactions or privacy protections
  • Commercial exploitation without proper legal review

Citation

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

@dataset{epstein_emails,
  title={Epstein Emails Dataset},
  author={Toon Beerten},
  year={2025},
  url={https://huggingface.co/datasets/[your-username]/epstein-phone-messages},
  note={Vision LLM-processed email dataset with live web application at epsteinsphone.org. Processing pipeline: https://github.com/Toon-nooT/epsteins-phone-reconstructed}
}

Remember: This dataset contains sensitive content from official government releases. Use responsibly and in accordance with all applicable laws, ethical guidelines, and the terms of the original House committee releases.

Dataset Version: 1.0
Last Updated: December 4, 2025
Processing Date: November 25, 2025