LightningRodLabs/war-gov-ufo-pursue-release-01

The WAR.gov UFO Files

161 public UFO/UAP files from WAR.gov in one easy table: official links, OCR text, image/video descriptions, transcripts when available, and conservative labels for browsing.

Source: WAR.gov PURSUE Release 01 at https://www.war.gov/UFO/

Important caveat: this is an unofficial research dataset. Columns beginning with original_ are copied from the official manifest. Columns beginning with inferred_ are machine-generated and are not official government determinations. Fields such as inferred_mentions_biologic_claim mean the source text appears to mention that kind of claim; they do not validate the claim.

Start Here

The default Hugging Face viewer is intentionally simple: one row per official file. Use war_gov_release_url to jump back to the WAR.gov release page, asset_url for the best human-facing row link, full_text for the complete OCR/transcript text, and media_description for image/video context.

Contents

• Dataset viewer: the default train split is the assets table, one row per official manifest row, including full_text when text/transcript is available.
• Downloadable archive: uap_pursue_tables.zip, with CSV and Parquet copies of assets, pages, chunks, media, cases, and field_definitions.

The official release has 161 manifest assets/files. The hosted viewer intentionally exposes only this 161-row assets table so the row count matches the number of official files/assets. Supporting page-level and chunk-level tables are downloadable in the archive for users who want citations or retrieval windows. For most users, assets.full_text is the simplest starting point.

Column order is intentional. assets starts with the asset ID, official title/type/agency, WAR.gov release URL, best human-facing asset URL, dates/location, complete full_text, media description, transcript, and high-level inferred fields. Lower-level OCR details, checksums, and source provenance are pushed to the right.

Quick Use

from datasets import load_dataset

ds = load_dataset("LightningRodLabs/war-gov-ufo-pursue-release-01", split="train")
print(ds.num_rows)  # 161 official assets
print(ds[0]["original_title"])
print(ds[0]["full_text"][:500])

Build Summary

• Assets: 161
• Pages with text records: 4,196
• Chunks: 5,782
• Media rows: 79
• Derived cases: 141
• Type counts: {"IMG": 14, "PDF": 119, "VID": 28}
• Agency counts: {"Department of State": 7, "Department of War": 82, "FBI": 57, "NASA": 15}

Processing

PDF text extraction uses embedded text when it has enough words on a page, otherwise Tesseract OCR is run on a rendered page image. Images are OCR'd with Tesseract, but low-confidence image OCR is not promoted into assets.full_text; for visual material, use media_description. Videos are linked rather than re-hosted.

Source provenance fields are prefixed so they are not confused with source/event metadata. The public assets table emphasizes human-facing links; raw video playback links such as MP4 and HLS .m3u8 URLs are kept in the downloadable media table, not in the front-door table. The public tables include official URLs, source-file byte sizes, content types, and SHA-256 checksums where applicable. Local build telemetry such as download status, local file paths, and processing timings is kept out of the public dataset.

Media rows include media_description generated from the source image or DVIDS thumbnail plus source context. The generated text is deliberately split into visible-image and source-context sentences. Audio/video transcripts are included in transcript only when an official caption/transcript source was available; otherwise transcript_status records that no DVIDS caption metadata was available.

Usefulness Notes

This is intended as a finding aid and ML-ready corpus, not as a claim-verification dataset. The source material includes redactions, scans, historical files, hearsay, and resolved/prosaic records mixed with unresolved records.