saturn

This dataset contains 105,484 rows of phoneme records linked to languages by glottocode, drawn from 8 different sources. Each row pairs a language identifier (2,177 unique glottocodes) with a phoneme (3,142 unique values, mostly 1-character IPA-like symbols) and a segment class. The segment_class breakdown is the most informative summary: consonants dominate at 72,282 rows, vowels account for 31,052, and tones only 2,150. Source coverage is uneven — 'ph' alone supplies about 34% of records, while the long tail (ra, spa, aa) is much smaller, which matters if you compare across sources. Glottocode frequency is also skewed: kham1282 and osse1243 each appear hundreds of times, suggesting some languages have far richer phoneme inventories recorded than others.

This is a sequential row identifier: every one of the 105484 rows has a unique value, running from min 1 to max 105484 with mean and median both at 52742.5 and skew 0.0. The perfectly uniform distribution and zero null/outlier rate confirm it carries no analytic signal beyond row ordering.

This column holds Glottolog language codes — fixed 8-character identifiers (len_mean 7.999, len_max 8) drawn from a vocabulary of 2,168 distinct codes across 105,484 rows. With a 97.94% duplicate rate and top codes like kham1282 (622) and osse1243 (483) repeating heavily, each code labels many rows rather than identifying them. The 2-character minimum length suggests a small number of malformed or truncated entries worth inspecting.

This column holds phoneme tokens — single-word strings averaging 1.5 characters with a max length of 11 and every row being one word. Despite 105,484 rows, only 3,142 unique values exist and 97.0% are duplicates, with single letters like 'm', 'i', 'k', 'j' dominating the top values. The small vocabulary (1,339 words) and tiny token sizes suggest these are IPA-like phonetic units rather than full words.

A 3-level categorical tag classifying segments as consonant, vowel, or tone, with no nulls across 105,484 rows. The distribution is heavily skewed: consonant dominates at 68.5%, vowel takes most of the rest, and tone is rare at only 2,150 occurrences. Entropy ratio of 0.64 confirms the imbalance.

Categorical provenance tag with 8 distinct codes (ph, ea, upsid, er, saphon, aa, spa, ra) marking which source each of the 105,484 rows came from. Distribution is fairly balanced for a source field — entropy ratio 0.899 and the top code 'ph' covers only 34.4% — suggesting the dataset is a merge of multiple comparably-sized corpora rather than one dominant source with minor supplements. No nulls, so every row is attributable.

Despite its name, created_at holds only 2 distinct timestamp values across 105,484 rows, both within one second of each other on 2026-01-06. This looks like a batch ingestion or load timestamp rather than per-record creation time, with 71.6% of rows sharing the dominant value. There is no temporal variation to exploit as a feature.