datamatters24/research-document-archive

Research Document Archive

Computational analysis of 234,630 declassified U.S. government documents across 7 archival collections. Output of a 13-step ML pipeline extracting OCR text, entities, topics, keywords, redactions, and semantic embeddings from 3.1 million pages.

Files

File	Rows	Description
documents.parquet	234,630	Document metadata: id, source_section, file_path, file_hash, total_pages
pages/<section>.parquet	3.1M	Per-page OCR text + 1536-dim sentence-transformer embeddings. Sharded by collection.
entities/<section>.parquet	31M	Named entities (spaCy en_core_web_lg): PERSON, ORG, DATE, GPE, FAC, LOC, NORP, EVENT. Sharded by collection.
document_topics.parquet	234,629	BART-large-MNLI zero-shot topic assignments (per doc, top-1 topic + probability)
document_features.parquet	1.17M	EAV feature table: redaction_summary, forensic_metadata, bertopic, sentiment, topic_distribution, exact_duplicate
document_keywords.parquet	3.5M	TF-IDF keywords (top 15 per document, unigrams + bigrams)
document_dates.parquet	234,630	Inferred document dates (regex + header parsing)
document_events.parquet	296K	Document-to-historical-event correlations (20 crisis events)
historical_events.parquet	20	Crisis event dictionary (event_name, date range, category, keywords)
entity_relationships.parquet	2.88M	Entity co-occurrence pairs with counts, distances, and sample documents

Collections

source_section	Documents	Source
cia_declassified	1,605	CIA Reading Room
cia_mkultra	1,936	MKULTRA release
cia_stargate	13,937	Stargate remote viewing program
doj_disclosures	—	DOJ public disclosures
house_resolutions	181,092	House.gov bill text (GovInfo API)
jfk_assassination	35,979	National Archives JFK release
lincoln_archives	21	Library of Congress

Loading

from datasets import load_dataset
docs = load_dataset("datamatters24/research-document-archive", data_files="documents.parquet")

# Or load a specific sharded table:
import pyarrow.parquet as pq
pages = pq.read_table("hf://datasets/datamatters24/research-document-archive/pages/cia_mkultra.parquet")

Methodology

• OCR: Tesseract + PyMuPDF
• NER: spaCy en_core_web_lg
• Topics: BART-large-MNLI (zero-shot) + BERTopic (unsupervised)
• Embeddings: sentence-transformers/all-MiniLM-L6-v2 (384-dim) and OpenAI text-embedding-3-small (1536-dim) per page
• Redaction detection: OpenCV contour analysis on PDF-rendered pages
• Entity relationships: page-window co-occurrence with distance weighting

Code: https://github.com/tedrubin80/Massivedata-Pull