| --- |
| license: other |
| task_categories: |
| - question-answering |
| - text-retrieval |
| language: |
| - en |
| tags: |
| - uap |
| - ufo |
| - government-records |
| - rag |
| - ocr |
| pretty_name: Official UAP Disclosure Corpus (OCR / native, page-level) |
| size_categories: |
| - 1K<n<10K |
| --- |
| |
| # Official UAP Disclosure Corpus |
|
|
| Page-level text of the recent **official U.S. government UAP (Unidentified Anomalous Phenomena) |
| disclosure**, extracted two ways and published side by side for comparison. 789 documents, |
| **7,288 pages**, drawn only from official releases — no third-party or speculative material. |
|
|
| ## Two extraction variants |
|
|
| | directory | how it was extracted | |
| |---|---| |
| | `data-ocr/` | **forced OCR** — the GLM-OCR vision model re-reads every page | |
| | `data-no-ocr/` | **native** text-layer extraction (xberg + pdf_oxide 0.3.74), OCR only as a scan fallback | |
| |
| Each variant ships a **global** file (`all.parquet`, `all.jsonl.gz`) covering the whole corpus, |
| **and one file per source dataset** (`<dataset>.parquet`, `<dataset>.jsonl.gz`), plus `SHA256SUMS`. |
| |
| **Schema** (per row = one page): `source` (dataset/relative-path), `page` (int), `text` (str), |
| `content_type` (`pdf` | `image`). |
|
|
| ## Source datasets |
|
|
| | dataset | records | docs | pages | |
| |---|---|--:|--:| |
| | `pursue-01` | war.gov PURSUE declassified release 01 | 129 | 4,179 | |
| | `pursue-02` | PURSUE release 02 (May 22 2026) | 6 | 128 | |
| | `pursue-03` | PURSUE release 03 (Jun 12 2026) | 24 | 1,302 | |
| | `pursue-04` | PURSUE release 04 (Jul 10 2026 — latest) | 13 | 554 | |
| | `aaro` | AARO Historical Record Report Vol. 1 (2024) | 1 | 63 | |
| | `odni-reports` | ODNI/DoD annual UAP reports (FY21 prelim, FY23, FY24) | 3 | 43 | |
| | `nasa` | NASA UAP Independent Study Team report (2023) | 1 | 36 | |
| | `nara-osd` / `nara-faa` / `nara-odni` / `nara-nrc` | NARA UAP electronic records | 612 | 983 | |
|
|
| ## OCR vs native — the headline finding |
|
|
| Forced OCR recovers **1.39× more text overall** (14.5M vs 10.4M characters), almost entirely from |
| the scanned PURSUE releases where native extraction sees little. On the born-digital reports (AARO, |
| ODNI, NASA, most NARA) native extraction is marginally richer and OCR is redundant. Because this |
| corpus is scan-dominated, `data-ocr` is the more complete version. See `reports/ocr-vs-no-ocr.md`. |
|
|
| ## Reports (`reports/`) |
|
|
| - `full-run-report.md` — corpus, pipeline, and results overview. |
| - `ocr-vs-no-ocr.md` — per-dataset extraction comparison. |
| - `defect-scan.md` — pdf_oxide structural defect scan (45 / 7,288 pages flagged). |
| - `ufo-qa-ocr.pdf` / `ufo-qa-no-ocr.pdf` — graded retrieval-augmented Q&A over each index |
| (Mistral-Large answerer, independent Opus 4.8 judge), with verbatim answers and cited sources. |
| data-ocr: faithfulness 0.96 / context precision 0.68; data-no-ocr: 0.94 / 0.78. |
| |
| ## Provenance |
| |
| Every document traces to an official government release. Acquisition method (the war.gov |
| `uap-data.csv` manifest and Wayback-Machine recovery for host-blocked files) is in |
| `reports/corpus-provenance.md`. One release-04 document (`CIA-UAP-D021`) is not archived anywhere |
| reachable and is noted as a known gap. |
| |
