svetfm/epstein-fbi-files

FBI Epstein Files - Embeddings Dataset

Document embeddings and OCR text from the FBI's release of Jeffrey Epstein-related files.

Dataset Structure

embeddings/
  all_embeddings.jsonl    # 236K chunks with 768-dim embeddings
ocr/
  all_ocr.jsonl          # Full OCR text for each document

Embedding Format

Each line in all_embeddings.jsonl is a JSON object:

{
  "id": "uuid",
  "bates_number": "EFTA00000001",
  "bates_range": "EFTA00000001-EFTA00000001",
  "source_volume": 1,
  "source_path": "VOL00001/...",
  "doc_type": "typed_memo",
  "ocr_confidence": 0.85,
  "ocr_engine": "textract",
  "page_number": 1,
  "total_pages": 3,
  "chunk_index": 0,
  "total_chunks": 5,
  "chunk_text": "...",
  "embedding": [0.123, -0.456, ...],  // 768 dimensions
  "ingested_at": 1703123456
}

Embedding Model

• Model: nomic-embed-text via Ollama
• Dimensions: 768
• Chunking: 1500 chars with 300 overlap

Statistics

• Documents: 8,150
• Unique with embeddings: 6,618
• Total chunks: 236,174
• Date range: 1990-2020

Usage

from datasets import load_dataset

# Load embeddings
ds = load_dataset("svetfm/epstein-fbi-files", data_files="embeddings/all_embeddings.jsonl")

# Access embeddings
for item in ds['train']:
    embedding = item['embedding']
    text = item['chunk_text']
    bates = item['bates_number']

Visualizations

Interactive visualizations available at: [GitHub Pages site - coming soon]

• UMAP embedding clusters
• Entity network graphs
• Timeline analysis
• Key player networks (Maxwell, Clinton, Trump, etc.)

License

Public domain government documents. Dataset compilation CC-BY-4.0.