| --- |
| license: mit |
| task_categories: |
| - text-classification |
| - token-classification |
| - text-mining |
| language: |
| - en |
| tags: |
| - government-documents |
| - nlp |
| - named-entity-recognition |
| - declassified |
| - jfk |
| - cia |
| - ocr |
| - document-analysis |
| size_categories: |
| - 100K<n<1M |
| --- |
| |
| # Research Document Archive |
|
|
| 234,630 declassified U.S. government documents processed through a 13-step ML pipeline. 3.2 million pages OCR'd, 31 million named entities extracted and linked, 288 topic clusters identified. |
|
|
| **Live platform:** [tanglewoodapp.com](https://tanglewoodapp.com) |
|
|
| ## Collections |
|
|
| | Collection | Documents | Pages | Size | |
| |---|---|---|---| |
| | House Resolutions | 181,092 | 2,719,832 | 34.2 GB | |
| | JFK Assassination Records | 35,979 | 241,860 | 22.5 GB | |
| | CIA Stargate Program | 13,937 | 100,056 | 5.4 GB | |
| | CIA MKUltra | 1,936 | 64,244 | 3.4 GB | |
| | CIA Declassified | 1,605 | 29,744 | 2.4 GB | |
| | Lincoln Archives | 21 | 9,330 | 962.9 MB | |
|
|
| ## ML Pipeline (13 Steps) |
|
|
| 1. Document ingestion and format normalization |
| 2. OCR with Tesseract + post-correction |
| 3. Classification stamp detection (SECRET, CONFIDENTIAL, UNCLASSIFIED, etc.) |
| 4. Redaction detection and boundary mapping |
| 5. Named entity recognition (people, organizations, locations, dates) |
| 6. Entity disambiguation and cross-document linking |
| 7. Relationship extraction |
| 8. Topic modeling (LDA + BERTopic) |
| 9. Timeline event extraction |
| 10. Network graph construction |
| 11. Sentiment and tone analysis |
| 12. Document similarity clustering |
| 13. Index building for search and retrieval |
|
|
| ## Classification Stamps Detected |
|
|
| | Stamp | Count | |
| |---|---| |
| | UNCLASSIFIED | 16,501 | |
| | SECRET | 13,736 | |
| | CLASSIFIED | 10,730 | |
| | EXEMPT | 6,739 | |
| | CONFIDENTIAL | 5,554 | |
| | RESTRICTED | 4,722 | |
|
|
| ## Key Statistics |
|
|
| - **31M** named entities extracted |
| - **2.9M** entity cross-document links |
| - **59,830** redactions detected and mapped |
| - **288** topic clusters identified |
| - **6** document collections spanning 1860s–2000s |
|
|
| ## Usage |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds = load_dataset("datamatters24/research-document-archive") |
| |
| # Filter by collection |
| jfk = ds.filter(lambda x: x["collection"] == "jfk_assassination") |
| |
| # Search by entity |
| cia_docs = ds.filter(lambda x: "CIA" in x["entities"]) |
| ``` |
|
|
| ## Data Sources |
|
|
| All documents are public record obtained from: |
| - National Archives (NARA) |
| - CIA FOIA Reading Room |
| - Congress.gov |
| - Library of Congress |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{rubin2026researcharchive, |
| author = {Rubin, Theodore}, |
| title = {Research Document Archive: ML Pipeline for Declassified U.S. Government Documents}, |
| year = {2026}, |
| publisher = {HuggingFace}, |
| url = {https://huggingface.co/datasets/datamatters24/research-document-archive} |
| } |
| ``` |
|
|
