saturn

Sighting is almost certainly a row identifier: every one of the 147890 values is unique, there are no nulls, and the distribution is essentially flat (skew -0.013, kurtosis -1.13) spanning 111 to 179773. The values look like sparse integer IDs rather than a measured quantity, since uniqueness equals row count and no outliers are flagged.

Stored as text but functionally a timestamp: values follow a 'YYYY-MM-DD HH:MM:SS Local' pattern with len_median 25 and word_median 3, and 'local' appears in 20000 of the top words. Of 147890 rows, 126264 are unique, yet 21626 duplicates (14.6%) cluster on holiday-evening times like 2014-07-04 22:00:00 and 2015-11-07 18:00:00. 299 rows contain only the bare token 'Local', indicating missing date/time fragments that need cleaning.

Short 'City, State, Country' location strings, with a mean length of 20.4 characters and 3.6 words. Heavy duplication (74.9%, 110,819 of 147,890 rows) reflects expected city repetition, with Phoenix, Seattle, and Las Vegas leading; the multilingual flag is misleading since 4,481 of detected rows are English and the non-English counts likely come from place names like 'Montréal' or 'São Paulo'. Across 37,070 unique values the dataset skews strongly to USA (17,880 word occurrences) and Canada.

Categorical descriptor of object shape across 147890 rows with 39 distinct values, dominated by 'Light' at 19.4% followed by 'Circle' (14367) and 'Triangle' (13086). Entropy ratio of 0.74 indicates a moderately spread distribution rather than one-class dominance, but the presence of both 'Other' (10062) and 'Unknown' (10021) in the top 5 signals substantial reporter ambiguity. Null rate is 4.3%, and several top categories ('Sphere', 'Orb', 'Circle') are likely semantic duplicates worth consolidating.

Free-text duration field expressing time spans like '5 minutes' or '30 seconds', averaging 2.07 words and 9.5 characters. Values are highly repetitive: 88.97% are duplicates with only 15,527 uniques across 147,890 rows, and the top word 'minutes' appears 8,115 times alongside 'seconds' and 'min'. The mix of units and abbreviations (e.g., 'min' vs 'minutes') signals inconsistent formatting that needs normalization, and 4.83% are null.

Numeric count of observers per record, with a tight typical range (q1=1, median=2, q3=2) but extreme tail reaching 20000 and a nonsensical min of -10. Skew (109.3) and kurtosis (14332.3) confirm a long upper tail, and 13.0% of values flag as outliers while 6.9% are zero and 4.5% are null. The negative minimum and 20000 maximum suggest data-entry errors or sentinel values worth scrubbing.

This is a report-date timestamp stored as a fixed 27-character string with a trailing 'Pacific' timezone label (e.g., '1999-11-16 00:00:00 Pacific'). All 147,890 values share identical length 27 and a 3-word structure, and 'pacific' appears in roughly 20,000 word occurrences, confirming a uniform date+tz format rather than free text. The 'multilingual' alert and small de/ru/zh language counts are spurious artifacts of language detection on numeric date strings, and a 7.7% duplicate rate (11,418 rows) reflects busy reporting days like 1999-11-16 (73 rows).

Posted holds 626 unique date-stamps (all at 00:00:00) across 147,890 rows with no nulls, so it's almost certainly the post date of each record. Distribution is highly diffuse — entropy ratio 0.93 and the top date 2020-06-25 accounts for only 1.24% (1,833 rows) — yet the top values span 1999 through 2023, suggesting a wide historical window rather than a recent dump.

This is a categorical multi-label feature describing observed object characteristics (e.g. "Lights on object", "Aura or haze around object", "Aircraft nearby"), stored as comma-separated tags rather than free text. Despite 147,890 rows, only 1,446 unique strings exist and 98.6% are duplicates, with "Lights on object" alone accounting for 29,911 entries. Note the 28.15% null rate and that the language detector only classified 3,527 rows as English — likely because most values are short tag strings.

Free-text summary field with 144,208 unique values across 147,890 rows, so it's effectively a per-record description rather than a category. Lengths swing wildly from 1 to 10,624 characters (mean 134.9, median 76), and the word distribution is dominated by English stopwords plus 'light' (4,676), suggesting narrative prose about sightings or observations. Readability is conversational (Flesch 67.3), with negligible URLs, emoji, or boilerplate, and only a 1.9% duplicate rate.

Free-text field containing medium-length English prose, averaging 949 characters and 181 words with a Flesch readability of 69.6, suggesting reasonably accessible writing. High lexical diversity (127,124 unique values across 147,890 rows, vocab 65,106) and frequent first-person tokens like 'i' and 'my' point to user-authored narratives such as reviews or stories. Worth flagging: 11.28% nulls, 3.1% exact duplicates (4,091 rows), and a max length of 64,550 characters indicating extreme outliers.

Free-text 'Location details' annotations, present on only ~6.9% of rows (null_rate 0.931) with 9,713 unique values across 147,890 records. Entries are short prose (len_mean 38.7, word_mean 6.9, Flesch 69.8) dominated by English function words like 'the', 'of', 'in', suggesting handwritten descriptions rather than structured location codes. Duplicate_rate 0.048 and 492 repeated strings hint at some recurring phrases, but the field is otherwise near-unique and sparsely populated.

This is a sparse explanatory label classifying observations into likely causes (Starlink, rocket, balloon, aircraft, planet/star, camera anomaly) paired with a confidence qualifier (Certain/Probable/Possible). It is almost entirely empty at 99.46% null, leaving only ~800 labelled rows spread across 58 categories with no dominant class (top_rate 0.0971, entropy_ratio 0.825). The compound 'Cause - Confidence' structure is worth splitting into two fields for cleaner analysis.