Datasets:

lilbeee
/

ufo

An end-to-end retrieval-augmented-generation build over the recent official Pentagon/IC UAP disclosure, run through the production pipeline (xberg + pdf_oxide 0.3.74 → Qwen3-Embedding-8B → LanceDB → parquet/jsonl; Mistral-Large-2411 answerer, independent Opus 4.8 judge, the #508 retrieval pipeline) and published as two extraction variants for comparison.

Start here

full-run-report.md — the master report: corpus, pipeline, the A/B extraction finding, defect scan, and graded Q&A results. Read this first.

Reports

ocr-vs-no-ocr.md — forced-OCR vs native extraction, page-aligned, per dataset.
defect-scan.md / .json — pdf_oxide structural defect scan over all 7,288 native pages.
ufo-qa-ocr.pdf / ufo-qa-no-ocr.pdf — the graded Q&A sessions with verbatim answers and cited sources (the no-ocr report includes an OCR-vs-native before/after page). HTML source alongside; machine-readable grades in qa-graded-ocr.json / qa-graded-no-ocr.json.
PLAN.md — the plan & storyboard.
corpus-provenance.md — how the corpus was sourced and staged (the Akamai-block workaround, the war.gov uap-data.csv manifest, Wayback recovery).

Corpus

789 documents, 7,288 pages. The official disclosure — PURSUE releases 01–04 (war.gov/UFO, including the Jul 10 2026 release 04) + NARA electronic records (OSD/FAA/ODNI/NRC) + AARO/ODNI/DoD reports (Historical Record Report Vol 1, FY21 prelim, FY23, FY24) + the NASA UAP study. Per-dataset page counts and sourcing in full-run-report.md and corpus-provenance.md.

Question sets

qa/questions.txt — 11 RAG-rigor questions (AARO counts/timeline, per-case object + evidence, sensor modalities, HRR reverse-engineering conclusions, sensitive-site sightings, cumulative counts…).
qa/questions_community.txt — 16 strictly community-driven questions, built by iterating public forum and press demand to convergence (non-human intelligence/biologics, crash-retrieval and reverse-engineering claims, named programs, the famous cases, nuclear-site incursions, transmedium behavior, whistleblower testimony, redactions, Blue Book vs AARO…).
2 refusal probes (in qa/capture_v31.py): in-domain-unsupported + out-of-corpus.

Two extraction variants (A/B)

data-ocr — forced GLM-OCR on every page.
data-no-ocr — native text-layer extraction (xberg + pdf_oxide), OCR only as a scan fallback.

Each is published with a global all.parquet/all.jsonl.gz and per-dataset <dataset>.parquet/.jsonl.gz, plus SHA256SUMS.

Where every artifact lives

HF beeberg/ufo (https://huggingface.co/datasets/beeberg/ufo): data-ocr/ and data-no-ocr/ (global + per-dataset parquet/jsonl + checksums), and the reports under reports/.
RunPod volume lilbee-corpus-us (US-MO-1): the staged raw corpus at /workspace/ufo and the LanceDB index snapshots (ufo-index-ocr.tar, ufo-index-no-ocr.tar).
This branch (ufo-corpus): the reports (md + committed PDF), graded JSON, and the full staging
- ingest + QA harness.

Reproducibility

infra/ufo_download.sh — idempotent corpus staging (war.gov CSV manifest, Wayback id_ recovery, snapshot pinned past the latest release, skip guards).
infra/ingest_ufo.sh — two-variant ingest (force-OCR / native + fallback), global + per-dataset export.
infra/ingest_ufo_r4.sh — the append run that folded in release 04 without re-OCR'ing the corpus.
infra/ufo_qa.sh + qa/capture_v31.py — serve Mistral-Large on each index and capture both question sets.
infra/run_ufo.sh / run_ufo_r4.sh — the launchers (land pod → ingest → QA → power off).