alex-zhang42/ufo-pursue-open-atlas

The U.S. Department of War PURSUE Release 01 UFO / UAP declassification, re-extracted into cleaned Markdown with inline image captions — every photograph, sketch, rubber stamp, and handwritten margin note becomes an *Image: ...* block in the same text stream. Interleaved image-text data derived from 80 years of declassified gov documents, with per-page JPEG renders, interactive 3D atlas, and a fully client-side keyword + AI hybrid search. All CC0.
_{161 records · 4,153 pages · ~2 GB image data · 2026-05-11 · v0.2}

Get started · Search · Screenshots · Schema · Why VLM · Considerations · Cite

---

What's in here

Artifact	Size	What it is
corpus.jsonl	14 MB / 4,153 rec	Per-page Markdown with inline *Image: ...* descriptions for every photograph, sketch, stamp, handwritten note. One JSON object per line. The form shipped on this GitHub repo.
pages/train-*-of-00005.parquet (5 shards, on HF only)	~2.0 GB	Same per-record fields as corpus.jsonl plus an image column carrying the 200 DPI JPEG bytes inline as the leftmost column. Built locally by scripts/build_parquet_shards.py; published on the 🤗 Hub as the pages config (gitignored on GitHub — too large for git). load_dataset(..., "pages") returns decoded PIL.Image objects directly.
mimo_processed/	25 MB	Per-page Markdown sources (one directory per PDF, with _meta.json provenance).
image_audit/	8 MB	Per-page audit JSON — mimo judge scores plus GPT-mini outputs.
web/	~140 MB	Plain HTML + JS for the 3D globe atlas, the side-by-side dataset viewer, and the v0.2 keyword + AI search. Includes precomputed search index (web/search_index/, ~13 MB) and 400 px page thumbnails (web/thumbs/, ~125 MB) so first-load works offline. No build step, runs on any static server.
pipeline/ + scripts/	source	Full extract → audit → fallback → render → build → validate pipeline, plus build_search_index.py / build_thumbnails.py for the v0.2 search artifacts and search_cli.py for offline retrieval QA.
corrections.json	20 KB	Every metadata correction (4 location fixes, 1 date fix, 56 inferred N/A dates) with rationale.

License: CC0 1.0. Source documents are works of the U.S. federal government, already public domain under 17 U.S.C. §105.

---

Get started

Load the corpus

Two configs — pick by what you need:

from datasets import load_dataset

# text + metadata (default; ~14 MB, fast)
ds = load_dataset("alex-zhang42/ufo-pursue-open-atlas", split="train")
print(ds[0]["text"][:200])

# text + metadata + decoded page image (~2 GB across 5 parquet shards)
ds = load_dataset("alex-zhang42/ufo-pursue-open-atlas", "pages", split="train")
print(ds[0]["text"][:200])
print(ds[0]["image"])  # PIL.Image, already decoded

Field map is in Schema below; full reference in schema.md.

Run the atlas + viewer + search locally

The web/ directory is plain HTML + JS — no build, no Node, no install.

git clone https://github.com/AlexZhangji/ufo-pursue-open-atlas
cd ufo-pursue-open-atlas
python3 -m http.server 8000 --directory web/

Then open:

• http://localhost:8000/ — 3D globe atlas (all 161 records by agency × time × location)
• http://localhost:8000/dataset.html — side-by-side viewer (every PDF page next to its VLM Markdown)
• http://localhost:8000/search.html — keyword + AI hybrid search (v0.2; details below)

Or browse online

Hosted mirror of the same web/ directory at:

• https://ufo.gpt2077.com/ — atlas
• https://ufo.gpt2077.com/dataset.html — side-by-side viewer
• https://ufo.gpt2077.com/search.html — search (v0.2)

Walk per-PDF (if you cloned the repo)

import json, pathlib
for meta in pathlib.Path("mimo_processed").glob("*/_meta.json"):
    for page_md in meta.parent.glob("page_*.md"):
        text = page_md.read_text(encoding="utf-8")
        ...

---

What it looks like

Side-by-side viewer (run locally, or hosted mirror). Every page of every PDF on the left, cleaned VLM Markdown on the right — every claim in the corpus is one click from its source page.

3D globe atlas (run locally, or hosted mirror). 161 records colored by agency, jittered so dense locations stay individually clickable, time-linked to a brushable 80-year histogram.

---

What the Markdown layer preserves

Every page is VLM-extracted into clean Markdown. Every photograph, sketch, diagram, chart, rubber stamp, and handwritten margin note becomes an inline *Image: <factual description>* block in the same text stream. The Markdown carries:

• Structural text — headings, body, numbered + bulleted lists, block quotes
• Tables — form fields, distribution lists, signature blocks rendered as proper Markdown tables
• Classification banners — ## UNCLASSIFIED / ## CONFIDENTIAL / ## SECRET as Markdown headings
• Image content — inline *Image: <factual description>* blocks
• Rubber + ink stamps — quoted verbatim and tagged inline
• Handwritten annotations — preserved as *[handwritten: ...]* italics where they appear in the page flow
• Redactions — black-bar redactions noted as [REDACTED] with surrounding context
• Margin annotations + page numbers — kept as italic asides

86.6% of the 4,153 source PDF pages have zero native text — they are image-only scans. For those pages this Markdown layer is the only source of searchable text. See Why not just use the embedded PDF text? below for the empirical sweep.

1947 FBI archival page mounting a newspaper clipping ("Priest Finds 'Whirring' Disc..."), with a handwritten FBI distribution list down the right margin. The Markdown layer preserves: full article body transcribed verbatim under a "More About Saucers" heading; the distribution list rendered as a bulleted list; date stamps captured inline; and an `*Image:*` block describing the page surface. Native PDF text on this page: 0 chars.

---

Search interface (v0.2)

Live at ufo.gpt2077.com/search.html. Stamp-style header with the same wax-seal emblem on this README, query bar with a row of click-to-fill examples, filter chips annotated with bucket counts (DOW 1,346 · FBI 2,417 · ...), and a multi-column grid of result cards. Default query ships as newspaper clipping flying saucer so first-load shows the corpus's most visually distinctive pages.

Hybrid keyword + AI search across every page of every PDF. Both retrievers run in the browser — no server, no telemetry, query string never leaves your machine.

Layer	What it does	Tech
Keyword (BM25)	Exact + prefix + fuzzy match over title, text, image_tags, agency, incident_location with field-weighted boosts. Catches named entities ("Apollo 17", "Section 4").	MiniSearch
AI (dense vector)	384-dim semantic match using bge-small-en-v1.5 embeddings. Query is encoded in-browser via transformers.js; page vectors are precomputed offline. Catches paraphrased queries ("how fast did the object move" → 1350 MPH page with zero word overlap).	transformers.js
Fusion	Per-query min-max normalize both score lists, weighted sum (default keyword 0.6 / AI 0.4). Slider in Advanced to retune live.	—

Other niceties: page thumbnails inline · markdown-stripped readable snippets with query-term highlighting · ?q=... URL state · agency / era / record-type filter chips with bucket counts · service-worker offline cache (~13 MB index + ~25 MB ONNX model, both cached after first load).

Rebuild the search index

Only needed if you change corpus.jsonl or want to retune the embedding model.

uv sync                                  # installs fastembed + numpy + the rest
python scripts/build_thumbnails.py       # → web/thumbs/  (~125 MB · 4153 JPEGs)
python scripts/build_search_index.py     # → web/search_index/  (~13 MB total)

Total runtime ~6 min on a 2024 laptop CPU (no GPU needed; fastembed uses ONNX). Sanity-check retrieval quality before deploying:

python scripts/search_cli.py "how fast did the object move"
python scripts/search_batch.py    # batch eval of 12 candidate UI queries

Retrieval honesty

• Tested at page granularity — median page is 369 tokens, comfortably within BGE's 512-token window; ~15% of long pages (NASA debriefings, FBI thick volumes) get tail-truncated by the encoder.
• BM25 here is bag-of-words: phrasal queries like "intelligence assessment" don't score for adjacency. Dense retrieval covers most of that gap; if you need true phrase matching, the easiest path is bigram tokenization on the MiniSearch side (~30 min).
• English only — the corpus and embedding model are both English.

---

Schema (one record per page)

{
  "pdf_stem": "065_HS1-834228961_62-HQ-83894_Section_4",
  "page_num": 191,
  "source_url": "https://www.war.gov/medialink/ufo/release_1/...",
  "sha256": "022b27...",
  "agency": "FBI",
  "year": 1949,
  "year_inferred": false,
  "title": "FBI 62-HQ-83894 Section 4",
  "incident_location": "Virginia, USA",
  "text": "## UNCLASSIFIED\n\n*Image: A clipped newspaper article titled...*\n\n...",
  "vlm_model": "mimo-v2.5",
  "vlm_prompt_version": "v3.1-uap-archive",
  "image_tag_source": "gpt-5.4-mini-2026",
  "image_tag_audit_score": 1
}

image_tag_source is mimo-v2.5 for the majority of pages with image tags and gpt-5.4-mini-2026 for the 515 pages where mimo's first pass showed inconsistencies (263 in the original audit batch + 252 in a second batch of pages mimo had mis-categorized as stamp_or_label_only / typed_text_only despite genuine visual content). See DATA_CARD.md for full audit methodology.

Data fields

Full reference + nullability rules in schema.md.

Field	Type	What it is
record_id / pdf_stem	string	Stable per-record key. Maps 1:1 to mimo_processed/<stem>/.
page_num / total_pages	int	1-indexed page within the source PDF.
text	string	Full Markdown of the page, including inline *Image: ...* blocks.
text_chars	int	len(text).
image_tags	array[string]	Just the inside of each *Image: ...* block, in document order.
image_tag_source	string	mimo-v2.5 or gpt-5.4-mini-2026 (515 pages re-described after audit).
image_tag_audit_score / _audit_categories	int / array, nullable	Mimo-judge consistency score (0-3) and strict-extract categories for re-described pages; null otherwise.
source_url / sha256 / file_size_bytes	string / string / int	Direct war.gov download URL + integrity hash + size of the source PDF.
agency	string	Department of War / Department of State / FBI / NASA.
record_type	string	mission_report / cable / case_file / photo / imagery / report / summary / transcript / other.
title	string	Cleaned record title from the master CSV.
incident_location / _corrected	string|null / bool	Cleaned location, plus a flag if we corrected the source CSV.
incident_date_iso / year / year_inferred / _corrected	string|null / int|null / bool / bool	ISO date if known, year (recovered from filename for the 56 N/A records), and provenance flags.
description_blurb / dvids_video_id	string|null / string|null	Source CSV description blurb and paired DVIDS video ID where applicable.
vlm_model / vlm_prompt_version / vlm_dpi / extraction_completed_at	string / string / int / string	Provenance of the Markdown extraction.
page_image_path	string|null	Repo-relative path to the rendered page JPEG. Present in the text config; the pages config replaces this with an inline image column of decoded bytes.
image	datasets.Image()	Only in the pages config. Decoded JPEG returned as PIL.Image.
page_image_format / render_dpi / render_max_dim / render_jpeg_quality / render_version	string / int / int / int / string, all nullable	Render parameters. render_version is bumped when render params change.

---

Considerations for using the data

Full limitation list in DATA_CARD.md §7. Five things worth knowing before you use it:

1. Not a verbatim OCR substitute. text is a VLM re-rendering of the source PDF, not the embedded text layer. For any quotation, citation, or evidentiary use, verify against the source page — the side-by-side viewer makes that one click.
2. No human in the loop. Both extraction passes are LLM. Subject-matter verification is left to you.
3. 86.6% of the corpus is image-only scanned PDF with zero native text. For those pages the VLM is the only source of text. See "Why not just use the embedded PDF text?" below.
4. Image-tag audit is partial. 515 of 1,236 image-tag pages were re-described by gpt-5.4-mini; the remaining 721 keep their original mimo-v2.5 description. Treat image_tag_source as a coverage flag if you are running statistics on image content.
5. PII in source documents. Source records are already public under 17 USC §105, but they do name living people (State Department cables name foreign politicians, FBI case files name civilian witnesses). Apply normal care when quoting individuals.

---

Why not just use the embedded PDF text?

A reasonable question — modern PDF readers extract text "for free", so a VLM extraction pipeline at first looks redundant. The empirical answer, swept across all 4,153 pages of the v0.1 corpus with pymupdf.page.get_text():

Native text (fitz / pymupdf)	Page count	Share
Zero chars (image-only scan, no embedded text)	3,595	86.6%
<50 chars (typically just stamp text or page numbers)	35	0.8%
≥50 chars of real text	523	12.6%
└ of those, with detectable OCR glitches¹	45	~9% of 523

¹ Heuristic: COMFIDEMTIAL instead of CONFIDENTIAL, E.0. (zero) instead of E.O., broken fl ligatures fused into surrounding words, etc.

What this means in practice — three sample pages:

Source page	pymupdf native text	VLM Markdown
FBI 1947 newspaper clipping (62-HQ-83894 §4 p191)	0 chars	2,263 chars: full article transcribed + the surrounding archival page described as inline *Image:* ... block
1985 State Dept cable (digital PDF, 059uap00011 p1)	1,520 chars but reads "COMFIDEMTIAL", "E.0.: Unknown", "IUNCLASSI Fl ED"	1,530 chars, structured: # UNCLASSIFIED, ## CONFIDENTIAL, **MRN:** 01 MOSCOW 13169, **Date/DTG:** Oct 30, 2001
FBI 1947 typewriter letter with handwritten margins (62-HQ-83894 §1 p81)	0 chars	1,898 chars: full letter transcribed + handwritten annotations preserved as inline italics

The FBI 62-HQ-83894 case file alone is ~2,000 pages and every one returns empty from native PDF text extraction — there is no fallback for that subset. For the 12.6% of pages that do carry native text, the VLM pass adds classification banners as headings, form / cable headers as Markdown tables, inline image descriptions, and explicit handling of stamps and handwriting — none of which pymupdf.get_text() produces.

VLM-extracting the entire corpus uniformly (including the digital 13%) keeps schema and quality consistent. The tradeoff is that VLM output is non-verbatim — see Considerations #1.

---

Reproducibility

Pipeline runs end-to-end. Each stage is resumable.

# Setup
uv sync                                  # or: pip install -e .
cp .env.example .env                     # then fill in MIMO_TP_KEY + OPENAI_API_KEY

# Run
python3 scripts/download_from_csv.py uap-csv.csv         # 1. download (war.gov, browser headers required)
python3 pipeline/run_vlm_v3.py                           # 2. VLM Markdown extraction (mimo-v2.5)
python3 pipeline/recheck_images.py --concurrency 4       # 3. image-tag audit pass
python3 pipeline/run_fallback_mini.py --concurrency 4    # 4. gpt-5.4-mini re-describe flagged pages
python3 pipeline/apply_mini_replacements.py              # 5. apply re-descriptions
python3 scripts/render_pages.py -j 4                     # 6. 200 DPI page JPEGs
python3 scripts/build_corpus_jsonl.py                    # 7. build corpus.jsonl
python3 scripts/validate_corpus.py                       # 8. schema check
python3 scripts/build_parquet_shards.py                  # 9. pages-*.parquet shards (multimodal config)

Every artifact carries provenance — model, prompt version, DPI, source sha256, timestamps — in mimo_processed/<stem>/_meta.json.

---

Status

v0.1 covers PURSUE Release 01 (DoW, 2026-05-08): 161 records, 130 PDFs / images, 4,153 extracted pages, 1,236 image-tag pages of which 515 were GPT-mini re-described, all 4,153 pages rendered as 200 DPI JPEGs. Atlas live at https://ufo.gpt2077.com/.

Future PURSUE tranches (Release 02, 03, ...) will be appended on a rolling basis as DoW publishes them.

---

Citing

If you use this dataset, the suggested citation is:

@dataset{pursue_open_atlas_2026,
  title  = {PURSUE Open Atlas: A clean, machine-readable archive of
            U.S. Department of War declassified UAP records},
  author = {Zhang, Ji},
  year   = {2026},
  url    = {https://github.com/AlexZhangji/ufo-pursue-open-atlas},
  license = {CC0-1.0}
}

Or simply: PURSUE Open Atlas (CC0 1.0). Ji Zhang, 2026.