| --- |
| license: other |
| license_name: us-public-record |
| pretty_name: Epstein Files Corpus — OCR'd & Cleaned |
| task_categories: |
| - feature-extraction |
| - text-retrieval |
| tags: |
| - ocr |
| - legal |
| - court-records |
| - foia |
| - public-records |
| - epstein |
| - efta |
| size_categories: |
| - 10K<n<100K |
| --- |
| |
| # Epstein Files Corpus — OCR'd & Cleaned |
|
|
| Machine-readable text of the publicly released Jeffrey Epstein–related document |
| collections: the DOJ "EFTA" disclosures (Data Sets 1–12), House Oversight |
| releases, court records (Maxwell, Doe v. Epstein/Indyke, Florida v. Epstein, |
| USVI v. JPMorgan, and others), and FOIA productions (FBI, BOP, CBP, Florida). |
| All content is public-record material released by the U.S. Department of |
| Justice, federal/state courts, and FOIA respondents. |
|
|
| This is a **living dataset**: this snapshot publishes 32,172 documents — including the full House Oversight estate releases of Oct 17 (8,718 pages) and Nov 12 (22,903 pages); |
| roughly 33,000 source PDFs are downloaded and mid-OCR at snapshot time, |
| with over 1.3M catalog entries identified across 73 source datasets. |
|
|
| ## Repository layout |
|
|
| ``` |
| cleaned/<dataset>/<document_id>.md # 32,172 cleaned OCR markdown documents |
| (house-oversight-nov is sharded into per-1000 subdirs: cleaned/house-oversight-nov/<NN>/house-oversight-NNxxxx.md — Hugging Face caps directories at 10,000 files) |
| extracted/summaries/<dataset>/<id>.txt # 30,888 extractive summaries |
| extracted/embeddings.csv # chunk embeddings (document_id, chunk, vector) |
| extracted/entities.json # extracted entity index |
| catalogs/freeze-manifest-2026-08-27.json # acquisition freeze: per-dataset locator counts |
| catalogs/clean-manifest.json # cleaning-stage stats |
| catalogs/delineation-manifest.json # document-id -> source mapping, page/char counts |
| catalogs/summaries-manifest.json # summary index |
| catalogs/acquisition-plan-2026-08-27.md # how sources are enumerated & fetched |
| ``` |
|
|
| Datasets represented in this snapshot: `court-papers`, `efta-ds1/2/3/4/5/6/7/12`, |
| `flight-logs`, `foia-bop/cbp/fbi/florida`, `house-oversight-oct` (the House |
| Oversight October 17 estate release), `maxwell-proffer`, `memoranda`, |
| `phase-1`. |
|
|
| ## Provenance |
|
|
| - **DOJ EFTA releases** — [justice.gov/epstein/doj-disclosures](https://www.justice.gov/epstein/doj-disclosures), |
| Data Sets 1–12, EFTA-numbered PDFs. |
| - **Court records** — SDNY/SDFL and other dockets: *US v. Maxwell*, |
| *Doe v. Epstein*, *Doe v. Indyke*, *Florida v. Epstein* (2006/2008), |
| *USVI v. JPMorgan*, grand jury materials, proffer transcripts. |
| - **FOIA** — FBI, Bureau of Prisons, CBP, and Florida state releases. |
| - **House Oversight** — phase-1 document production, plus the October 17 |
| estate release (`house-oversight-oct`, 8,718 documents). |
|
|
| Each cleaned document traces to a source locator (URL + sha256) via the |
| freeze manifest and the delineation manifest. |
|
|
| ## OCR pipeline |
|
|
| 1. **Acquire** — source PDFs cataloged from official release pages |
| (`catalogs/freeze-manifest-*.json`), downloaded with sha256 verification, |
| mirrored to object storage (DigitalOcean Spaces, `epstein-index-files/raw/`). |
| 2. **OCR** — [marker-pdf](https://github.com/datalab-to/marker) converts PDFs |
| to markdown; low-confidence pages get a vision mop-up pass with |
| Qwen2.5-VL. |
| 3. **Delineate** — multi-document OCR outputs are split into per-document |
| files (9,269 documents in this snapshot). |
| 4. **Clean** — running headers, Bates stamps, disclaimers, and redaction |
| artifacts normalized (see `catalogs/clean-manifest.json`). |
| 5. **Extract** — extractive summaries, chunk embeddings, and entity index. |
|
|
| ## Quality caveats |
|
|
| This corpus is machine-OCR'd and **not human-proofread**. The 2026-08-15 |
| quality battery (11 sensor families, 63 sensors, 80 of 117 documents sampled) |
| produced **2,435 findings**; 100% of sampled documents had at least one |
| finding. Severity: 158 critical, 1,801 major, 275 minor, 201 info. The |
| dominant issue classes: |
|
|
| - `delineation.page-number-restart` — 1,066 findings (page numbering resets |
| across concatenated filings) |
| - `furniture.running-header` — 369 (residual running headers) |
| - `repetition.inline-loop` / `low-entropy-window` — 355 (OCR repetition loops) |
| - `legal.*` family — 280 (Bates-number gaps, stamp artifacts, transcript line |
| numbers, foreign-text regions) |
|
|
| Findings were measured against the *original* OCR tier; the `cleaned/` tier |
| published here has had the mechanical artifact classes (headers, stamps, |
| Bates furniture) addressed, but repetition-loop and delineation issues remain. |
| **Verify quotations against the source PDF before citing.** Reference-family |
| sensors (PDF text-layer cross-checks) could not run, so measured findings are |
| a floor, not a ceiling. |
|
|
| ## Raw PDFs |
|
|
| Source PDFs (~33 GB downloaded so far, of ~1.4 TB cataloged) are **not** |
| included in this snapshot. They currently live on DigitalOcean Spaces |
| (`epstein-index-files`, `raw/` prefix) with sha256 manifests in the freeze |
| catalog. They may be added here later as a separate `raw/` tree or a sibling |
| dataset repo once OCR coverage settles, to keep this repo fast to clone for |
| text-only consumers. |
|
|
| ## Update cadence |
|
|
| This repo is re-published incrementally as the OCR fleet processes the |
| backlog (~32k PDFs in flight). Expect snapshot commits as new document |
| batches clear delineation + cleaning; manifests are regenerated per batch. |
| Consumers should pin a commit SHA for reproducibility. |
|
|
| ## License |
|
|
| The underlying documents are U.S. public records (federal court filings, DOJ |
| releases, FOIA productions). The OCR/cleaning/extraction work product in this |
| repo is released to the public domain to the extent permitted; no rights are |
| asserted over the underlying documents. Note that released material may still |
| contain redactions, privacy-sensitive third-party information, and |
| sealing-related gaps as produced by the releasing agencies. |
|
|
