File size: 3,347 Bytes
000a710 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 |
This dataset is not raw UFO reports — it’s a *processed, enriched, semantically-clustered corpus* designed for large-scale analysis.
Below is the exact pipeline used.
---
# **🧠 1. Embeddings (Semantic Encoding)**
All reports were embedded using:
**Model:** `BAAI/bge-large-en-v1.5`
**Dimensionality:** 1024
Embeddings capture meaning (not keywords), allowing similar descriptions to cluster even with different phrasing, spelling, or vocabulary.
---
# **📉 2. Dimensionality Reduction (UMAP → 15D)**
High-dimensional vectors were reduced using:
**UMAP(n_components=15, metric='cosine')**
Reasons for UMAP-15:
* preserves local/global structure
* reduces noise
* improves cluster separation
* makes density-based clustering stable
---
# **📍 3. Density Clustering (HDBSCAN)**
Reports were grouped using:
**HDBSCAN(min_cluster_size≈tuned, min_samples≈tuned)**
Outputs include:
* `cluster_id` (−1 = noise)
* `prob` (cluster stability score)
* ~3.7k clusters
* ~20% noise
HDBSCAN discovers meaningful themes like:
* recurring object behaviors
* atmospheric misidentifications
* military-adjacent patterns
* long-term witness motif clusters
* hoax/storytelling clusters
* nonsensical/noise clusters
---
# **🔍 4. Sparse Retrieval (BM25) — Used for QA, Not in the Dataset**
A **BM25 index was built during preprocessing** to assist in quality control:
BM25 was used to:
* sanity-check embedding clusters
* inspect keyword cohesion
* identify outliers / mislabeled points
* verify that HDBSCAN clusters were semantically coherent
* detect keyword drift within large clusters
**Important:**
The *BM25 scores and index are **not included** in the final dataset.*
BM25 influenced the cleaning stage but is not part of the exported fields.
---
# **🌕 5. Sidecar Feature Enrichment**
Each record includes enriched metadata:
### **Moon illumination & altitude**
* `moon_illum`
* `moon_alt_deg`
Computed from timestamp + lat/lon.
### **Nearest airport (US/CA/GB accuracy strongest)**
* `nearest_airport_km`
* `nearest_airport_code`
Computed via geospatial lookup.
### **Weather bucket**
* `wx_bucket` (high-level NOAA-based label, imperfect)
### **Timestamp normalization**
* `ts` = Unix epoch (ms)
### **Source tagging**
* `src` indicates which Kaggle dataset the row came from.
---
# **📚 6. Canonical Output Format**
Each JSONL entry looks like:
```
{
"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": ...,
"ts": ...
}
```
---
# **📌 What This Dataset *Is***
✔️ A semantically-clustered UFO corpus
✔️ Enriched with astronomy + geospatial + weather sidecars
✔️ Cleaned, deduped, normalized
✔️ Built using modern ML (BGE+UMAP+HDBSCAN)
✔️ Ready for search, visualization, mapping, temporal analysis
✔️ Distributed without interpretation or claims
---
# **📌 What This Dataset *Is Not***
❌ Not a curated list of “important” sightings
❌ Not opinionated — no inferences built in
❌ Not a proof of anything
❌ Not filtered toward any outcome
❌ Does not include BM25 scores (BM25 was QA-only) |