saturn

This is a tornado event log with 70,022 rows and 13 columns covering dates, times, start/end coordinates, magnitudes, widths, fatalities, injuries, and U.S. state. Geographically it is a U.S.-centered dataset: starting longitudes average around -92.7 and latitudes around 37.1, with Texas (13.3% of records), Kansas, and Oklahoma leading the state counts. The severity fields are highly imbalanced — fatalities are 0 in 97.7% of events and injuries are 0 in 88.8% — so any analysis of harm should focus on the rare non-zero tail. Magnitude (mag) is a more usable categorical signal with 7 levels, dominated by 0 (46%) and 1 (34%). Note that the end-coordinate columns (elat, elon) are null in ~37.7% of rows, which matters if you plan to draw tornado tracks rather than just start points.

This is a date column stored as ISO-formatted text (YYYY-MM-DD), with every value exactly 10 characters long and one token. Across 70,022 rows there are only 12,639 unique dates and an 81.9% duplicate rate, so many records share dates — top value 2011-04-27 appears 207 times. The range spans at least 1974-04-03 to 2023-03-31, suggesting it was misclassified as text rather than a date type.

This column holds clock times stored as 8-character HH:MM:SS strings, with all 70,022 rows non-null and uniform length (len_min=len_max=8). Only 1,438 distinct values appear and 97.95% are duplicates, with afternoon/evening slots like 16:00:00 (978), 17:00:00 (971) and 15:00:00 (959) dominating — consistent with event start times rather than free text. It's mistyped as text: parse to a proper time type before use.

This is a US state code column with 53 distinct values, slightly more than the 50 states (likely including DC and territories like PR). TX dominates at 13.3% of 70,022 rows, followed by KS (4,474) and OK (4,221), giving a clear southern/plains skew rather than a uniform national distribution. Entropy ratio of 0.847 confirms the distribution is fairly spread but not flat, and there are no nulls.

`mag` is a low-cardinality categorical with 7 distinct values dominated by '0' (46% of 70022 rows) and decreasing counts through '1','2','3','4','5'. The ordered integer levels suggest a magnitude or severity code rather than a free category, and the presence of '-9' (1024 rows) is the standout signal — almost certainly a sentinel for missing/unknown that is not being counted as null.

Counts of injuries stored as strings, with 209 distinct values across 70,022 rows and no nulls. The distribution is severely zero-inflated: '0' accounts for 88.8% of records, and entropy ratio is just 0.123. Tail values like '1' through '10' decay quickly but 209 unique tokens suggests very long tails or non-integer entries worth inspecting.

This is a fatality count per event, stored as a categorical/string field with 50 distinct integer values and no nulls across 70,022 rows. The distribution is extremely imbalanced: '0' accounts for 97.72% of records, giving an entropy ratio of just 0.039, with '1' at 830 rows and a long thin tail (2 → 277, 3 → 134, down to single-digit counts at higher values). Despite being typed as categorical, the values are numeric and ordered, so the current encoding is likely a load artifact.

Numeric loss values stored as text strings, with all 70,022 entries being single tokens (one_word_rate 1.0) and lengths of 1-10 characters. The distribution is heavily concentrated: '0.0' alone accounts for 22,764 rows and the duplicate_rate is 0.985 across only 1,019 unique values. Mixed formatting is a hazard — '0' and '0.0' appear as separate tokens, so a naive cast will collapse them but string-based grouping won't.

Values range from 17.72 to 61.02 with a mean of 37.14 and median of 37.03, consistent with a starting latitude (slat) field in decimal degrees, likely covering the contiguous US. The distribution is essentially symmetric (skew 0.04, kurtosis -0.58) with a tight IQR of 7.74 and only 70 outliers (0.10%). No nulls and no zeros, and 16,016 unique values across 70,022 rows suggest repeated coordinates rather than free-form noise.

Values are negative decimal degrees ranging from -163.53 to -64.7151 with a median of -93.5, consistent with western-hemisphere longitudes (the 'slon' name suggests starting longitude). The distribution is mildly left-skewed (-0.32) and concentrated within an IQR of ~11.7 degrees around the central US longitude band. Only 951 outliers (1.36%) fall outside that range, and there are no nulls or zeros.

Almost certainly an event latitude in decimal degrees, with values spanning 17.72 to 61.02 — consistent with locations across North America. The distribution is roughly symmetric (skew 0.03, kurtosis -0.41) and centered near 37.26, with only 78 mild outliers. The standout concern is that 37.65% of rows are null, so coverage is partial.

This is almost certainly longitude (east coordinate), with values bounded between -163.53 and -64.7151 and a median of -92.47, consistent with points across North America. The distribution is moderately left-skewed (-0.60) with kurtosis 3.77 and a tight IQR of 11.26 around the median. Notably, 37.65% of rows are null, which will materially shrink any geo-based analysis.

Despite being typed as text, `len` holds short numeric tokens (length 3-8, one word each) like '0.1', '0.5', '1.0' — almost certainly a length or size measurement stored as strings. Values are highly concentrated: '0.1' alone covers 15,456 of 70,022 rows and the duplicate rate is 94.8% across only 3,663 unique tokens. The allcaps and one_word alerts are artefacts of numeric strings rather than real text signal.

wid is a categorical column with 419 distinct values, all numeric-looking strings like '10', '50', '100', '30' — almost certainly a width or weight parameter stored as text rather than a true category. The distribution is heavily concentrated on round numbers: '10' alone covers 20.6% of 70022 rows and the top ten values are all multiples of 5 or 25, giving an entropy ratio of 0.51. No nulls, but the string encoding of clearly numeric quanta is the surprise.