danger-ufo / PLANNING.md
sidsenthilkumar's picture
Upload folder using huggingface_hub (part 9)
b0c2370 verified
|
Raw
History Blame Contribute Delete
8.01 kB

UAP Consolidated Dataset — Planning Phase

What we're building

A single canonical UAP/anomaly dataset on HuggingFace, built from scratch. Not a mirror of existing datasets — a ground-up assembly with provenance discipline, uncertainty modeling, and adapter isolation as first-class concerns.

What's being abandoned

The old uap-spine project (SQLite + HTML viz on Spark). Its five architecture principles transfer cleanly. Its execution path (HF mirror → SQLite → static HTML) doesn't. This is a HuggingFace dataset, not a local database with a web frontend.

Context inventory (what's now in memory)

1. Existing processed data (tau/Ufo_data_clustered)

  • 327,009 records, JSONL, 142MB
  • Sourced from Kaggle NUFORC-derived CSVs
  • Fields: uid, t_utc, lat, lon, text, src, city, state, country, cluster_id, prob, moon_illum, moon_alt_deg, nearest_airport_km, nearest_airport_code, wx_bucket, reports_z
  • Pipeline: Cleaning → BGE-large-en-v1.5 embeddings → UMAP-15 → HDBSCAN → sidecar enrichment
  • Limitations:
    • Almost entirely NUFORC-lineage. No Blue Book, no AARO
    • Local→UTC timestamp conversion unverified (time_precision = hour at best)
    • Cluster labels are text-similarity only, not event categories
    • Likely duplicates across Kaggle mirror lineages
    • Shape and duration fields dropped during cleaning
    • Astronomical sidecars (moon_*) depend on unverified timestamps — do not trust

2. Raw source CSVs (from tau source/)

  • nuforc_reports.csv (116MB), scrubbed.csv (14MB), ufo_sightings_scrubbed.csv (14MB), ufo_sighting_data.csv (14MB), complete.csv (15MB), data.csv (43MB), weatherHistory.csv (16MB), airports.csv (8.7MB), runways.csv (3.0MB), airport-frequencies.csv (1.2MB), definitions_and_sources.csv (57KB)
  • All Kaggle-derived

3. Raw CSV from downloads

  • ufo_sightings_1900_2024.csv at /mnt/shared/MyFiles/Downloads/ufo-dataset-downloads/ — alternate source, needs inspection for overlap/redundancy

4. Global dataset landscape (from overview document)

  • Civilian: NUFORC, MUFON, Kaggle mirrors
  • US Government: AARO (imagery, case resolutions), PURSUE (rolling declassification tranches, unredacted kinematics), NARA RG 615/341 (Blue Book, Condon, Roswell, Presidential libraries), FBI/CIA vaults
  • International: GEIPAN (France, weirdness/consistency scoring), SEFAA (Chile, aviation focus), UK National Archives, LAC (Canada), NAA (Australia)
  • Academic: Galileo Project (Harvard, custom calibrated observatories), SCU on Zenodo (morphological analysis, operational presence, Aguadilla), NASA Independent Study
  • ML hazard: acronym collisions with "UAP" (Universal Adversarial Perturbations) and "UFO" (Urban Flood Observations, Unified Fine-grained Perception)

5. Architecture principles (from spine project — still correct)

  1. Spine as canonical schema. Every source adapts into one fixed record shape. Source-specific extras go in extra JSON blob.
  2. Append-only, provenance immutable. source, source_record_id, original_text, fetched_at locked at write time. No overwrites. Reports are separate evidentiary artifacts, not views of one truth.
  3. Uncertainty as first-class fields. Every fuzzy field gets a companion: time_precision, geocode_method, geocode_confidence, location_uncertainty_km.
  4. Environmental context = on-demand join. Weather, RF, flight, seismic, geomagnetic NOT backfilled into records. Joined at query time against event_time + lat/lon.
  5. Adapter isolation. Each adapter reads one source. Does not dedupe. Does not cross-reference. Resolution is downstream.

6. Canonical spine schema (from spine project — the frozen contract)

field type nullable notes
id text (uuid4) no assigned at ingestion
event_time text (ISO8601 UTC) yes best-known event time
time_precision text no second / minute / hour / day / month / year / unknown
latitude real yes
longitude real yes
location_uncertainty_km real yes
geocode_method text no source_provided / mirror_geocoded / city_centroid / unknown
geocode_confidence text no high / medium / low / unknown
source text no our source tag
source_record_id text no id in source's own namespace
original_text text no verbatim narrative from source
fetched_at text (ISO8601 UTC) no when we pulled it
extra text (JSON) yes source-specific extras

Constraints: PK on id, indexes on event_time, source, (latitude, longitude). No unique constraint on (source, source_record_id) — dedup is downstream.

7. Source ingestion plan (from spine project — adapted for HF)

Order Source Tag What it contributes
1 NUFORC-derived Kaggle data hf_mirror_cjc0013 327k records, fastest path to working pipeline
2 NUFORC raw nuforc_raw Shape/duration fields, authentic submission metadata
3 Project Blue Book bluebook Disposition field (official USAF classifications)
4 GEIPAN (France) geipan Scientifically vetted, weirdness/consistency scored
5+ AARO, SEFAA, Galileo, SCU tbd Document-centric / sensor-validated / academic

8. Old spine project's open forks (still relevant)

  • Repo name: uap-spine was placeholder. Needs a real name.
  • Source order: HF mirror first (fastest path), then whichever stresses untested architecture. NUFORC-raw stresses geocoding/timezone. Blue Book stresses source-specific field handling.
  • disposition field: stays in extra, not promoted to spine (avoids schema pressure from future sources with their own labels)
  • AARO adapter shape: document corpus → OCR/extraction → intermediate events → spine. Two-stage pipeline, not one monolithic adapter.
  • Cross-source identity resolution: downstream queryable relation, not a mutation on the spine. event_identity table mapping spine record clusters to canonical event ids.
  • Storage: HF datasets as primary distribution. Parquet format. Local SQLite for development/testing.

What needs decisions (planning phase)

A. Dataset name

uap-spine was placeholder. Considerations: short, doesn't overpromise, doesn't bind to a single source. Open.

B. HuggingFace format

Parquet is standard for HF datasets. Could also ship JSONL for line-by-line access. Default: Parquet as primary, JSONL as convenience mirror.

C. Adapter implementation language

Python with datasets library. Each adapter is a script that reads source → writes Parquet shards to data/ directory → datasets.load_dataset() reads them.

D. First adapter scope

Option A: Clean re-ingestion of the HF mirror (adapt existing JSONL → spine schema with uncertainty fields populated honestly, drop the untrusted sidecars). Option B: NUFORC-raw first (get shape/duration fields, do geocoding properly). Option C: Both in sequence, starting with the mirror for speed.

E. What to do with the existing clustered data

The cjc0013/Ufo_data_clustered dataset already exists on HF. This project is a separate dataset with a different contract (spine schema + provenance). The existing data is a source, not the destination. We map it in, we don't extend it.

F. The 1900-2024 CSV

The ufo_sightings_1900_2024.csv at /mnt/shared/MyFiles/Downloads/ufo-dataset-downloads/ is unexplored. Need to determine: source lineage, overlap with other sources, field completeness, whether it adds anything NUFORC-derived sources don't.

Immediate next step

The planning deliverable is answers to A–F above from Sid (name, format, first-adapter scope, posture on existing data). Then: adapter #1 build.


Context consolidated 2026-07-30. Sources: tau/Ufo_data_clustered, old spine docs, UFO Datasets Overview and Links.md, freshly built ufo-consolidated repo.