Epstein Files: Structured Data Exports

Structured data exports from the forensic analysis of the 218GB DOJ Jeffrey Epstein file release (all 12 datasets + House Oversight Estate + FBI Vault: 1,385,916 documents, 2,771,231 pages).

I have threaded through these databases into a searchable visual interface, with an AI-assistant, at https://epstein-data.com

You can also install my full Claude-powered Epstein Investigator setup yourself, by following the directions for the: desktop install, or CLI version.

Note the latest release, v5.1, at: https://github.com/rhowardstone/Epstein-research-data/releases/tag/v5.1

Results repo: Epstein-research — 165+ forensic investigation reports with DOJ source citations.

What's Here

Knowledge Graph (Curated)

File	Records	Description
knowledge_graph_entities.json	606	Curated entities: people, shell companies, organizations, properties, aircraft, locations. Each entry includes aliases, metadata (occupation, legal status, mention counts), and entity type.
knowledge_graph_relationships.json	2,302	Relationships between entities with types (traveled_with, associated_with, owned_by, victim_of, etc.), weights, date ranges, and source/target entity names.

Note on knowledge graph: The knowledge graph was curated during the initial investigation phases and does not include NER (Named Entity Recognition) run against the full OCR corpus. It covers the most frequently-referenced and manually-verified entities. For comprehensive name extraction from the full text, see extracted_entities_filtered.json below or query the full databases directly.

Person Registry

File	Records	Description
persons_registry.json	1,614	Unified person registry merged from 9 sources: epstein-pipeline (1,195), knowledge-graph (285), la-rana-chicana (237), Wikipedia Epstein files list (45), Bondi PEP letter Feb 2026 (19), jmail.world (9), corpus-investigation (2), khanna-massie-2026 (2), and doj-release-2026 (1). Each entry includes name, aliases, category (political/business/academic/staff/financial/legal/media/other), search terms, and source attribution.

Note: This registry is broader than the knowledge graph — it includes every named individual identified across all investigation phases, congressional disclosures, and cross-referenced sources. Categories reflect the person's primary role relative to the Epstein case, not an accusation. Many entries (e.g., Bondi PEP letter names) appear in the files only in incidental contexts such as news clippings or tips.

Entity Extraction (Filtered from 107K raw)

File	Records	Description
extracted_entities_filtered.json	8,085	Filtered entity extractions: 3,881 names (appearing in 2+ documents), 2,238 phone numbers, 1,489 amounts, 357 emails, 116 organizations. Each entry includes the EFTA document numbers where it appears.
extracted_names_multi_doc.csv	3,881	Names appearing in multiple EFTA documents with document counts and sample EFTA references. CSV format for easy browsing.

Note on quality: The raw extraction table contains 107,422 entities, many of which are OCR artifacts from redacted/degraded documents. The filtered exports remove garbled text and require multi-document co-occurrence for names.

Image Catalog

File	Records	Description
image_catalog.csv.gz	38,955	Complete image catalog (gzipped). Fields: id, image_name, efta_number, page_number, people, text_content, objects, setting, activity, notable, analyzed_at.
image_catalog_notable.json.gz	38,864	Images with people or notable content identified (gzipped JSON). Truncated fields for manageable size.

Document Summaries

File	Records	Description
document_summary.csv.gz	519,438	Per-document redaction summary for every EFTA document (gzipped). Fields: efta_number, total_redactions, bad_redactions, proper_redactions, has_recoverable_text, dataset_source.

Reconstructed Pages (High Interest)

File	Records	Description
reconstructed_pages_high_interest.json.gz	39,588	Pages where hidden text was recovered from under redactions (gzipped JSON). Fields include efta_number, page_number, num_fragments, reconstructed_text, interest_score, and names_found. Higher interest scores indicate more substantive recovered content.

EFTA-to-DOJ URL Mapping

File	Description
efta_dataset_mapping.csv	EFTA number ranges for each of the 12 DOJ datasets, with URL templates.
efta_dataset_mapping.json	Same mapping in JSON format for programmatic use.

URL Pattern: https://www.justice.gov/epstein/files/DataSet%20{N}/EFTA{XXXXXXXX}.pdf

Dataset	EFTA Start	EFTA End
1	00000001	00003158
2	00003159	00003857
3	00003858	00005586
4	00005705	00008320
5	00008409	00008528
6	00008529	00008998
7	00009016	00009664
8	00009676	00039023
9	00039025	01262781
10	01262782	02205654
11	02205655	02730264
12	02730265	02858497

Note: EFTA numbers are assigned per page, not per document. A multi-page document consumes consecutive EFTA numbers — e.g., EFTA00008320 (89 pages) covers Bates numbers 00008320–00008408, and Dataset 5 begins at EFTA00008409. There are no gaps between datasets 1–11; every apparent gap is accounted for by multi-page documents at dataset boundaries. Dataset 12 (the post-release expansion) contains internal gaps totaling ~100K unassigned EFTA numbers — these likely represent documents not yet released or reserved number ranges.

DOJ Document Removal Audit

The doj_audit/ directory documents an audit of the DOJ Epstein Library that identified documents removed or altered after the initial public release.

File	Records	Description
doj_audit/CONFIRMED_REMOVED.csv	67,784	Documents confirmed removed from the DOJ website (returning HTTP 404). Fields: efta, justice_gov_url, dataset, pages, scan_content_length, scan_last_modified.
doj_audit/FLAGGED_documents.csv	96,112	All flagged documents with DOJ URLs and dataset info.
doj_audit/FLAGGED_documents_details.csv	102,223	Flagged documents with detailed metadata including status, category, document type, priority score, confidence, and text preview.
doj_audit/SIZE_MISMATCHES.csv	23,989	Documents where the file size on the DOJ server differs from the originally ingested version, suggesting post-release modification.
doj_audit/sample_verification_results.csv	500	Statistical sample verification of flagged 404s using browser-based checks.

See the full report: DOJ Document Removal Audit

Alteration Analysis

The alteration_analysis/ directory contains analysis of documents where content was altered between versions of the DOJ release.

File	Records	Description
alteration_analysis/classified_alterations.csv	21,803	Documents with classified alteration types (CONTENT_REDUCTION, EMAILS_REMOVED, etc.) with sensitivity ratings and LLM-generated reasoning.
alteration_analysis/removed_entities_export.csv	146,209	Entities (names, accounts, phone numbers) that were removed from documents between versions, with corpus hit counts and classification.

The full alteration database (42,782 files tracked, 212,730 change units) is available as alteration_results.db.gz in the v5.1 release. See the full report: DOJ Document Alteration Forensics

Recovered Corrupted PDFs

The recovered_corrupted_pdfs/ directory contains text recovered from 5 corrupted PDF documents in the DOJ release through forensic byte-level carving:

EFTA00593870, EFTA00597207, EFTA00645624, EFTA01175426, EFTA01220934

Each subdirectory contains the cleaned extracted text from the corresponding corrupted PDF. See the full report: Corrupted PDF Forensics

Full Database Downloads

Source databases are split across releases:

• v5.0 — Updated full text corpus (DS12 expansion, March 2026)
• v5.1 — Alteration analysis database
• v4.0 — All other databases

Database	Release	Compressed	Uncompressed	Contents
full_text_corpus.db.gz	v5.0	2.3GB (split)	6.3GB	1,385,916 documents, 2,771,231 pages with full text, FTS5 search index. All 12 EFTA datasets + House Oversight Estate (DS99) + FBI Vault (DS98) + native spreadsheets + recovered EFTAs. Download both .part_aa and .part_ab and concatenate: cat full_text_corpus.db.gz.part_* > full_text_corpus.db.gz
concordance_complete.db.gz	v5.1	137MB	696MB	1,385,519 documents, 2,788,208 pages — concordance cross-reference with email threads, folder inventory, production metadata
alteration_results.db.gz	v5.1	183MB	8.2GB	212,730 change units with diff text, pixel-diff results, LLM classification
redaction_analysis_v2.db.gz	v4.0	166MB	971MB	2.59M redaction records, 849K document summaries, 39K reconstructed pages, 107K extracted entities
redaction_analysis_ds10.db.gz	v4.0	87MB	532MB	Dataset 10 deep analysis (EFTA01262782-02205654)
image_analysis.db.gz	v4.0	64MB	389MB	38,955 images with AI-generated descriptions
ocr_database.db.gz	v4.0	25MB	68MB	OCR extraction data
transcripts.db.gz	v4.0	1.7MB	4.8MB	1,628 media file entries, 435 with speech content, 189,982 words (faster-whisper large-v3)
knowledge_graph.db	v4.0	764KB	764KB	Curated entities and relationships (uncompressed SQLite). Updated entity/relationship JSON files with 606 entities and 2,302 relationships are in the repo directly.
communications.db.gz	v4.0	—	—	Email thread analysis
prosecutorial_query_graph.db	v4.0	2.5MB	2.5MB	Subpoena analysis: riders, returns, clause fulfillment, investigative gaps

Total: ~3.0GB compressed / ~17GB uncompressed

# Download and decompress the full text corpus (split into 2 parts)
wget https://github.com/rhowardstone/Epstein-research-data/releases/download/v5.0/full_text_corpus.db.gz.part_aa
wget https://github.com/rhowardstone/Epstein-research-data/releases/download/v5.0/full_text_corpus.db.gz.part_ab
cat full_text_corpus.db.gz.part_* > full_text_corpus.db.gz
gunzip full_text_corpus.db.gz

# Search the full text corpus
sqlite3 full_text_corpus.db "SELECT efta_number, page_number, substr(text_content, 1, 200) FROM pages WHERE text_content LIKE '%Leon Black%' LIMIT 10;"

# Search redacted content
sqlite3 redaction_analysis_v2.db "SELECT efta_number, page_number, substr(hidden_text, 1, 300) FROM redactions WHERE hidden_text LIKE '%TERM%' AND length(hidden_text) > 20 LIMIT 20;"

Processing Tools (Replication Pipeline)

The tools/ directory contains all Python scripts used to build the databases from raw PDFs. Use these to replicate the analysis, extend it with new data, or adapt for your own pipeline.

All tools auto-detect the data directory (no path editing needed). They check the EPSTEIN_DATA_DIR environment variable first, then look relative to the script and current working directory. If auto-detection fails: export EPSTEIN_DATA_DIR=/path/to/your/data

Core Pipeline

Tool	Description
tools/ingest_house_estate.py	Ingests House Oversight Estate documents (Concordance format, OCR with configurable workers)
tools/ingest_spreadsheets.py	Ingests native XLS/XLSX/CSV files into full_text_corpus.db
tools/transcribe_media.py	GPU transcription of audio/video using faster-whisper large-v3
tools/prescreen_media.py	Pre-screens media files to classify and skip surveillance footage
tools/redaction_detector_v2.py	Spatial redaction analysis: finds black rectangles, extracts underlying text
tools/build_person_registry.py	Builds unified person registry from 6 sources
tools/build_knowledge_graph.py	Constructs entity relationship graph
tools/build_native_files_catalog.py	Generates NATIVE_FILES_CATALOG.csv

Search & Analysis

Tool	Description
tools/person_search.py	FTS5 cross-reference search with co-occurrence analysis and CSV export
tools/congressional_scorer.py	Scores documents by redacted-name density for congressional reading room prioritization
tools/generate_gov_reports.py	Searches corpus for current government officials
tools/search_judicial.py	Searches corpus for federal judges
tools/extract_subpoena_riders.py	Extracts and catalogs subpoena rider documents

Data Integrity

Tool	Description
tools/find_missing_efta.py	Gap detection across EFTA numbering
tools/recover_missing_efta.py	Recovers missing EFTAs from DOJ server or forensic carving
tools/run_post_ingestion_pipeline.sh	Chains all post-ingestion steps (transcription, registry, catalog)

Data Acquisition

The raw PDFs can be obtained from:

Source	URL	Contents
DOJ Epstein Library	justice.gov/epstein	Datasets 1-12 (individual PDFs). Bulk downloads removed Feb 6, 2026.
Archive.org DS9	full.tar.bz2	103.6 GiB. Largest single dataset.
Archive.org DS11	DataSet 11.zip	25.6 GiB. 267,651 PDFs.
Archive.org DS1-5	combined-all-epstein-files	First 5 datasets combined.
House Oversight	oversight.house.gov	Estate documents, DOJ-provided records, photo releases.

See COMMUNITY_PLATFORMS.md in the research repo for a full directory of 78+ community tools and mirrors.

Integration Notes

For developers building tools on top of this data:

• EFTA numbers are the universal key. Every document in the DOJ release has one.
• The efta_dataset_mapping files let you resolve any EFTA number to a DOJ PDF URL.
• Entity efta_numbers arrays give you cross-references: "this person appears in these documents."
• Knowledge graph weight on relationships indicates strength of connection (higher = more documented).
• Image image_name format is EFTA{number}_p{page}_i{index}_{hash}.png — parse EFTA number and page from the filename.
• No inter-dataset gaps (DS1–DS11): EFTA numbers are per-page, so a multi-page terminal document in each dataset consumes the EFTA numbers up to the next dataset's start. DS12 (post-release expansion) has internal gaps.

License

This is analysis of public government records released under the Epstein Files Transparency Act (Public Law 118-299). The underlying documents are U.S. government works. This structured data is released into the public domain.

Contact

Please open an issue if you find any problems! We will respond promptly.