KyivNotKiev/toponym-adoption-data
Toponym adoption data English-language adoption of Ukrainian vs Russian toponym transliterations (Kyiv/Kiev, Chornobyl/Chernobyl, ...), 2010-2026. Layout Four stages, each a function of the previous. Only raw is expensive; everything downstream is a recompute. file what it is <source>_raw.parquet exactly what the provider returned, nothing dropped <source>_processed.parquet cleaned and regex-matched; only records containing a spelling… See the full description on the dataset page: https://huggingface.co/datasets/KyivNotKiev/toponym-adoption-data.
Toponym adoption data
English-language adoption of Ukrainian vs Russian toponym transliterations (Kyiv/Kiev, Chornobyl/Chernobyl, ...), 2010-2026.
Layout
Four stages, each a function of the previous. Only raw is expensive; everything downstream is a recompute.
raw deliberately contains material the study excludes -- videos about Vladimir Putin rather than Vladimir the Great, taxonomy papers citing the botanist T. Borsch, articles about Odessa, Texas. They are what the provider returned, and they are the evidence that the filtering did something. Without them the filtering cannot be audited.
Columns in _processed and pairs/
record_id, pair_slug, source, doc_id, url, date, title, text, ua_hits, ru_hits, variant, match_context, text_hash
There is no verified column. ua_hits and ru_hits are word-boundary match counts and variant is derived from them; match_context is the surrounding text so the match can be checked by eye. The evidence is the data.
pairs/ carries source because it is needed at evaluation time -- the first question about any cluster is whether it merely rediscovered the source. It must not be used as a training feature.
Coverage is uneven, on purpose
Not every pair has every source. Wikipedia, Trends and Ngrams produce counts rather than documents and so appear in no pairs/ file. YouTube census collection is partial. The per-pair files show what each pair actually has rather than implying uniformity.