saturn

This dataset contains 76,475 rows of linguistic feature observations, all sourced from WALS (World Atlas of Language Structures). Each row pairs a language (2,659 unique language IDs) with one of 192 typological features (e.g., 'Order of Object and Verb') and a categorical value encoded as both a short code (value_name) and a small integer (value). The most common features cluster around word order and negation, with 'Order of Object and Verb' being the top feature at 1,518 rows. Worth a closer look: the `value` column is highly skewed (skew 3.49, kurtosis 16.4) with ~3.2% outliers reaching up to 28, suggesting most features have only a few possible values but a handful have many categories. The `source` column is constant ('WALS') and can be ignored as a variable.

This column holds 2-3 character language codes (ISO 639-style: eng, fre, ger, rus...), with len_mean 2.996 and one_word_rate 1.0. With 76,475 rows but only 2,659 unique values and a 96.5% duplicate_rate, it behaves as a categorical key rather than free text. The top codes are remarkably balanced (eng 159, fre 158, ger 157), suggesting a curated multilingual catalogue rather than organic usage.

feature_id is a categorical code with 192 distinct values across 76,475 rows and no nulls, with codes like '83A', '82A', '143E' suggesting a numeric prefix plus letter suffix scheme. The distribution is remarkably flat: entropy ratio 0.937, top value '83A' covers only 1.98% of rows, and four '143' suffixes ('143A','143E','143F','143G') tie at exactly 1325 occurrences, hinting at a structured co-occurrence rather than random sampling.

This column names linguistic typology features, almost certainly WALS-style feature labels (e.g., 'Order of Object and Verb', 'Order of Subject and Verb'). With 192 distinct values across 76,475 rows and a top rate of just 1.98%, the distribution is remarkably flat (entropy ratio 0.94), suggesting each feature is sampled across many languages rather than a few features dominating. There are no nulls and no obvious duplicates among the top values.

A small-integer count or rating feature with only 28 distinct values ranging from 1 to 28, mean 2.85 and median 2. The distribution is heavily right-skewed (skew 3.49, kurtosis 16.36), with 2,469 outliers (3.23%) stretching the tail to 28 while the IQR sits tight at 1–4. No nulls or zeros, so every row carries a positive count.

This column holds short alphanumeric codes (e.g. '143G-4', '82A-1') — uppercase, single-token, 4-7 characters long, with 1.0 one_word_rate and 1.0 allcaps_rate. Across 76,475 rows there are only 1,139 distinct values and a 0.985 duplicate_rate, so it behaves as a categorical key rather than free text. The pattern (digits + letter + dash + digit) suggests a typology or feature-value identifier drawn from a fixed vocabulary of 911 tokens.

This column records the data origin and holds the constant value "WALS" across all 76,475 rows. With cardinality of 1 and entropy of 0.0, it carries no information for modelling and merely tags the dataset's provenance.