Datasetpublic

offshore-studio/croatian-profanity-list

Croatian Profanity List Content warning: vulgar, sexual, insulting, and hate-related terms in Croatian and related South Slavic forms. Intended for content moderation, search filtering, and NLP safety — not harassment. Open MIT wordlists + severity scores (0–5) for Croatian (HR) with regional overlap for BA / RS / ME. Resource URL GitHub (source of truth) https://github.com/offshore-studio/croatian-profanity-list Live collector (Psovke)… See the full description on the dataset page: https://huggingface.co/datasets/offshore-studio/croatian-profanity-list.

sourceHugging Facemitupdated 3d agoView on Hugging Face

0likes68downloads

Dataset Card

Croatian Profanity List

Content warning: vulgar, sexual, insulting, and hate-related terms in Croatian and related South Slavic forms. Intended for content moderation, search filtering, and NLP safety — not harassment.

Open MIT wordlists + severity scores (0–5) for Croatian (HR) with regional overlap for BA / RS / ME.

Resource	URL
GitHub (source of truth)	https://github.com/offshore-studio/croatian-profanity-list
Live collector (Psovke)	https://offshore.studio/psovke/
Export API	https://offshore.studio/psovke/api/export

Configs

Config	File	Rows (approx.)	Description
`terms` (default)	`data/terms.csv`	~268k	All surface forms with severity + flags
`lemmas`	`data/lemmas.csv`	~16k	Seed lemmas from `lists/hr.txt`
`phrases`	`data/phrases.csv`	~2k	Multi-word phrases

Raw originals also ship under data/raw/ (hr.txt, hr-severity.json, regex, …).

Columns (`terms`)

Column	Meaning
`term`	Surface form / lemma / phrase
`severity`	`2–5` when scored (empty if unscored)
`is_lemma`	`1` if in `hr.txt`
`is_phrase`	`1` if in `hr-phrases.txt`
`is_normalized`	`1` if in `hr-normalized.txt`

Severity scale

Score	Meaning
0–1	Mild / usually ignore (rarely exported)
2	Mild insult
3	Strong insult / vulgar
4	Heavy sexual / slur-adjacent
5	Most severe

Quick load

python

from datasets import load_dataset

ds = load_dataset("offshore-studio/croatian-profanity-list", "terms")
print(ds["train"][0])
# {'term': '...', 'severity': 4, 'is_lemma': 0, 'is_phrase': 0, 'is_normalized': 1}

Filter strong terms only:

python

strong = ds["train"].filter(lambda r: r["severity"] is not None and r["severity"] >= 4)

Intended use

—UGC / search-query moderation
—Training and evaluating NLP filters
—Research on South Slavic offensive language
—Crowdsourced discovery via Psovke

False positives

See the GitHub repo tests/hr_false_positive_tests.txt (e.g. kurir dostava, serija filmova). Matchers should use word boundaries / allowlists.

Citation

bibtex

@misc{croatian-profanity-list,
  title        = {Croatian Profanity List},
  author       = {offshore.studio},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/offshore-studio/croatian-profanity-list}},
  note         = {Also https://github.com/offshore-studio/croatian-profanity-list}
}

License

MIT — same as the GitHub repository.