README.md

---
license: mit
language:
  - en
task_categories:
  - text-classification
  - time-series-forecasting
size_categories:
  - 100K<n<1M
pretty_name: "Danger UFO Dataset"
tags:
  - ufo
  - uap
  - anomalous-phenomena
  - nuforc
  - provenance
  - geospatial
configs:
  - config_name: default
    data_files: "data/danger-ufo-dataset/*.parquet"
---

# Danger UFO Dataset

A canonical, provenance-disciplined dataset of 327,009 UFO/UAP sighting reports. Built from
scratch with first-class uncertainty modeling, adapter isolation, and append-only provenance —
not a mirror of existing datasets.

## Design principles

1. **Spine as canonical schema.** Every source maps into one fixed record shape.
   Source-specific extras live in a JSON `extra` blob — the spine never changes to
   accommodate a new source.
2. **Append-only, provenance immutable.** `source`, `source_record_id`, `original_text`,
   and `fetched_at` are locked at write time. No overwrites. Reports are separate
   evidentiary artifacts, not views of one truth to reconcile.
3. **Uncertainty as first-class fields.** Every fuzzy field has a companion:
   `time_precision`, `geocode_method`, `geocode_confidence`, `location_uncertainty_km`.
   A `latitude` with no companion says nothing about whether it came from GPS or
   geocoding "somewhere near Denver."
4. **Environmental context = on-demand join.** Weather, RF, flight, seismic data are
   NOT backfilled into records. Joined at query time against `event_time` + `lat/lon`.
5. **Adapter isolation.** Each adapter reads exactly one source. No dedup, no
   cross-referencing. Resolution is downstream.

## Quick start

```python
from datasets import load_dataset

ds = load_dataset("sidkrishna/danger-ufo-dataset")
print(f"{len(ds['train']):,} records")

# Or read Parquet directly
import pandas as pd
df = pd.read_parquet("data/danger-ufo-dataset/")
```

## Schema

| Field | Type | Nullable | Description |
|---|---|---|---|
| `id` | string (uuid4) | no | Assigned at ingestion |
| `event_time` | string (ISO8601 UTC) | yes | Best-known event time |
| `time_precision` | string | no | One of: `second`, `minute`, `hour`, `day`, `month`, `year`, `unknown` |
| `latitude` | float64 | yes | Decimal degrees |
| `longitude` | float64 | yes | Decimal degrees |
| `location_uncertainty_km` | float64 | yes | Radius of uncertainty |
| `geocode_method` | string | no | `source_provided`, `mirror_geocoded`, `city_centroid`, `unknown` |
| `geocode_confidence` | string | no | `high`, `medium`, `low`, `unknown` |
| `source` | string | no | Our source tag (e.g. `hf_mirror_cjc0013`) |
| `source_record_id` | string | no | ID in the source's own namespace |
| `original_text` | string | no | Verbatim sighting narrative |
| `fetched_at` | string (ISO8601 UTC) | no | When we pulled the source data |
| `extra` | string (JSON) | yes | Source-specific fields that don't fit the spine |

### Companion field semantics

- **`time_precision`**: How precisely we know the event time. Current source uses `hour`
  because the upstream local→UTC conversion is unverified.
- **`geocode_method`**: How lat/lon was derived. Current source uses `mirror_geocoded`
  because coordinates come from upstream Kaggle mirror geocoding, not original GPS.
- **`geocode_confidence`**: Coarse trust level. Current source uses `medium` — Kaggle
  mirrors have quality variation.
- **`location_uncertainty_km`**: Currently null for this source (the mirror didn't
  preserve this data). Will be populated for sources that provide it.

## Current sources

| Source tag | Description | Records |
|---|---|---|
| `hf_mirror_cjc0013` | NUFORC-derived Kaggle aggregations, cleaned and clustered (via `cjc0013/Ufo_data_clustered`) | 327,009 |

## Planned sources

| Order | Source | What it adds |
|---|---|---|
| 2 | NUFORC raw | Shape/duration fields, authentic submission metadata, proper geocoding |
| 3 | Project Blue Book | USAF disposition classifications (1947–1969) |
| 4 | GEIPAN (France) | Scientifically vetted cases with weirdness/consistency scoring |
| 5+ | AARO, SEFAA, Galileo Project, SCU | Sensor-validated, document-centric, and academic data |

## What's in `extra`

For `hf_mirror_cjc0013` records, the `extra` JSON blob contains all source fields that
don't map to the spine: `cluster_id`, `prob`, `moon_illum`, `moon_alt_deg`,
`nearest_airport_km`, `nearest_airport_code`, `wx_bucket`, `city`, `state`, `country`,
and `original_src` (the Kaggle source tag). HDBSCAN cluster labels reflect text
similarity only — not verified event categories.

## Known limitations

- **Timestamp precision:** `time_precision = hour` for all current records. Upstream
  local→UTC conversion is unverified. Do not treat `event_time` as more precise than
  an hour boundary.
- **Geographic bias:** 97%+ of records are from US/Canada. This is a reporting bias,
  not a physical distribution.
- **No deduplication:** Records may appear multiple times across different Kaggle
  mirror lineages. Cross-source dedup is a downstream operation.
- **No shape/duration:** These fields were dropped during upstream cleaning. They exist
  in raw NUFORC but haven't been ingested yet (planned as source #2).
- **Cluster labels are not event categories:** `cluster_id` and `prob` in `extra` come
  from text-similarity grouping (BGE → UMAP → HDBSCAN). They surface linguistic themes,
  not verified event types.

## Format

- **Parquet** (primary): 7 shards, 59 MB total, zstd compression
- **JSONL** (convenience): 1 file, 264 MB, for line-by-line access

## License

Source data was public on Kaggle. This cleaned, schema-normalized version is released
under MIT for research and educational use.

## Build

```bash
python3 adapters/hf_mirror.py
```

Reads from `data/raw/ufo_data_clustered.jsonl` (the upstream HF mirror JSONL),
writes Parquet shards to `data/danger-ufo-dataset/` and JSONL to
`data/danger-ufo-dataset.jsonl`.