| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| tags: |
| - government |
| - documents |
| - OCR |
| - parse |
| - reducto |
| pretty_name: UFOCR |
| size_categories: |
| - n<1K |
| --- |
| |
| # UFOCR: Declassified UFO/UAP Documents Parsed with Reducto |
|
|
| UFOCR is an open dataset of the FBI's and U.S. Department of War's declassified records on UFOs, UAPs (unidentified aerial phenomena), and extraterrestrial investigations — fully parsed into clean, structured, LLM-ready text using [Reducto](https://reducto.ai). |
|
|
| The original government archives are a notoriously difficult OCR challenge: scanned typewriter pages, faded carbon copies, handwritten margin notes, redacted blocks, rotated scans, multi-column memos, and dense tables. UFOCR is what those documents look like after being run through a modern document parsing pipeline built for exactly this kind of mess. |
|
|
| About the Dataset |
| - **Source**: Public declassification releases from the [FBI Vault](https://vault.fbi.gov/UFO) and the [Department of War UFO Archive](https://www.war.gov/UFO/) |
| - **Content**: All documents and images from both collections, parsed end-to-end |
| - **Format**: Structured text + layout metadata, ready for retrieval-augmented generation (RAG), fine-tuning, search indexing, or analysis |
| - **License: CC BY 4.0** — free to use commercially with attribution |
| - **Language**: English |
|
|
| ## Why this dataset exists |
| Declassified government archives are some of the highest-friction documents on the public internet. The raw PDFs are technically "available," but practically unreadable at scale — you can't grep across thousands of scanned typewriter pages, and naive OCR mangles tables, redactions, and handwriting badly enough to make downstream LLM work unreliable. |
|
|
| We released UFOCR to (1) make a culturally interesting corpus actually usable for AI research and journalism, and (2) demonstrate what high-accuracy document parsing looks like on real-world worst-case inputs. |
|
|
| ## How it was built: parsing with Reducto |
| [Reducto](https://reducto.ai) is a document ingestion and parsing API used by AI teams at companies like Harvey, Scale AI, Vanta, and Toast to turn complex PDFs, scans, spreadsheets, and slides into LLM-ready structured data. |
|
|
| Every page of the FBI and DoW UFO archives was processed through Reducto's parsing pipeline, which combines: |
| - Layout-aware computer vision to detect regions, tables, figures, and reading order on each page |
| - Vision-language models (VLMs) that interpret each region in context — linking labels to values, reading handwriting, and preserving table structure |
| - Agentic OCR that reviews its own outputs in real time and corrects mistakes before producing the final result |
|
|
| The result is parsed text that preserves the structure of the original documents — including tables, multi-column layouts, handwritten annotations, and bounding box metadata — instead of the flat, garbled string you typically get from off-the-shelf OCR. |
|
|
| *If you see any mistakes or documents with incorrect parsing, reach out to the team at support@reducto.ai to flag.* |
|
|
| ### Use cases |
| - **RAG over declassified archives** — build chatbots and research tools grounded in primary government sources |
| - **Fine-tuning and pretraining** — high-quality OCR'd English text from a unique domain |
| - **Historical and journalistic research** — full-text search across decades of UFO/UAP investigations |
| - **Benchmarking OCR and document AI systems** — the FBI/DoW corpus is a brutal stress test for any parsing pipeline |
|
|
| ### Try Reducto on your own documents |
| If you have your own difficult documents — financial filings, medical records, insurance forms, legal contracts, scanned archives, multilingual scans, handwritten notes — Reducto handles them through a single API. The platform supports PDFs, images, spreadsheets, and slides across 100+ languages, with SOC 2 and HIPAA compliance and self-hosted deployment options for regulated industries. |
|
|
| Get started: |
| - Explore the product: [reducto.ai](https://reducto.ai) |
| - Talk to our team: [reducto.ai/contact](https://reducto.ai/contact) |
|
|
