UADetox: Ukrainian toxic comments with neutral and literary rewrites

UADetox is a parallel dataset of 48,414 toxic Ukrainian social-media comments. Every comment has two rewrites that keep its meaning and stance but drop the rudeness:

• detoxified: a neutral, minimal-edit rewrite in modern standard Ukrainian;
• literary_detoxified: the same point in the register of classic Ukrainian literature, in one of four sub-styles (after Нечуй-Левицький, Франко, Леся Українка, Котляревський).

It is meant for training and evaluating detoxification and style-transfer models for Ukrainian.

Project page with the model and results: uadetox.pages.dev. Live demo: Housto4/uadetox. Model: Housto4/uadetox-e2b.

from datasets import load_dataset

ds = load_dataset("Housto4/uadetox")   # train / validation / test
row = ds["train"][0]
print(row["original"], "->", row["literary_detoxified"])

Example

original	detoxified	literary_detoxified (lesya)
А тепер почніть вже, курва, діяти А не пиздіти, сидячи в кріслах	А тепер почніть уже діяти, а не говорити, сидячи в кріслах.	Почніть же, панове, діяти, а не говорити з м'яких крісел! Гідність — у чині, не в слові.

Terms of use

The comments were collected from public Telegram channels (plus a small set of LLM-written synthetic comments and a public CC BY 4.0 Telegram corpus, see below). The dataset is released for non-commercial research only: training and evaluating models that detect or reduce toxicity, and studying style transfer. Do not use it to generate, amplify or target abusive content, and do not try to identify the people who wrote the comments. Telegram's terms of service apply to the underlying content. Rows from aisbergpublicorganization/telegram-news-ua-dataset are additionally subject to its CC BY 4.0 licence and must be attributed to it.

How the data was built

1. Comments. Most comments were scraped from public Ukrainian Telegram channels in two rounds:

   • telegram_scraped: comments under posts of public channels in 12 topic categories, kept in their natural distribution;
   • telegram_scraped_2: a second scrape (68 channels, 2022–2026) selected to cover rare words the first round lacked: war slang, mocking nicknames, surzhyk, gaming and internet slang.

   Candidates were pre-filtered by an automatic toxicity score and an LLM screen for toxicity, Ukrainian language and usability. User mentions, links and phone numbers were replaced with [user], [url] and [phone]. The curated part adds comments from the public corpus aisbergpublicorganization/telegram-news-ua-dataset (CC BY 4.0) and synthetic, LLM-written comments in the style of Facebook, Instagram and YouTube, generated from a persona × topic × toxicity-type × severity grid.

2. Rewrites. Large language models wrote both rewrites for each comment, following a written specification (keep the meaning, the target and the stance; remove insults, obscenity and dehumanising language; use correct modern Ukrainian without russisms; for the literary rewrite, use the assigned sub-style). For slang-heavy comments the writer received glosses of the slang terms and looked up unknown nicknames.

3. Checks. Each row passed rule checks (obscenity and slur lexicons, russisms, Russian letters, stray Latin script, refusal or meta text, length) and independent LLM judges for remaining offensiveness and for meaning preservation. Failing rows were sent back with the reasons and rewritten, then checked again; rows that still failed were left out.

4. Audits. Blind 1–5 audits of stratified 400-row samples scored meaning, fluency, literary style and offensiveness after each major round, and targeted fixes followed (for example, thin literary rewrites were strengthened and group generalisations removed).

5. Gold audit. An LLM judge reviewed every row of all three splits for meaning loss, meaning reversal, fluency errors, leftover toxicity and weak literary style. Flagged rows were rewritten by a fixer and verified by a separate checker; rows that could not be fixed were dropped. Across the three splits 7,667 flagged rows were rewritten and 6,779 flagged rows were reviewed and kept unchanged.

6. Dictionary fix. Words in the rewrites that are missing from a Ukrainian spelling dictionary were classified as valid (archaic, dialect, literary, loanwords) or errors (Russian words, wrong forms, misspellings, calques). Rows with an error were minimally edited and every edit was independently checked (263 rows, about 0.5%, changed).

7. Deduplication and splits. No comment appears twice, exactly or as a near-duplicate (character 5-gram MinHash, Jaccard ≥ 0.8), within or across splits.

Fields

field	type	description
id	string	row id
original	string	the toxic comment, with personal data masked ([user], [url], [phone]); otherwise kept as written, including typos, emoji and surzhyk
detoxified	string	neutral rewrite
literary_detoxified	string	literary rewrite
literary_style	string	sub-style of the literary rewrite: nechuy, franko, lesya, kotlyarevsky
rewrite_type	string	paraphrase; light_edit (near-neutral input, small edit); brush_off (pure abuse with no content, rewritten as a polite refusal with the same intent)
severity	int (0–5)	severity of the original comment on the rubric below
toxicity_types	list of strings	profanity, insult, identity_attack, threat, sexual, dehumanization, implicit
platform	string	telegram, or facebook / instagram / youtube for synthetic comments
origin	string	real or synthetic
subset	string	telegram_scraped, telegram_scraped_2 or curated
source_dataset	string	telegram-scraped, telegram-scraped-2, synthetic or aisbergpublicorganization/telegram-news-ua-dataset

Severity

severity is an ordinal level of how offensive the original comment is, defined by a written rubric. Two independent LLM annotators rated every comment on the rubric; the level is the band of their mean rating.

level	rubric	share
0	neutral; possibly negative opinion, no rudeness	2.2%
1	rude tone, mild slang, disparaging things rather than people	4.3%
2	sarcasm or condescension at a person, implicit insult, a single obscene word used as an intensifier	30.2%
3	several obscenities, or a direct but mild insult at a person or group, dehumanising war labels	35.4%
4	strong obscene insult at a person, identity slur, crude sexual remark	23.5%
5	threats or calls for violence, sexual insults of relatives, wishing death, stacked slurs	4.4%

Agreement of the two annotators on 47,919 comments: same level 63%, within one level 96%, quadratic-weighted Cohen's κ 0.78. Treat severity as an ordinal label, not a probability, and adjacent levels as close. Comments at level 0–1 are near-neutral inputs; most of them carry rewrite_type = light_edit.

Splits

The split is 80/10/10 over the whole corpus, grouped by thread (post, or channel-day for the second scrape) so that no thread or parent post appears in more than one split, and stratified by subset, platform and severity.

split	rows	telegram_scraped	telegram_scraped_2	curated
train	38,878	17,190	15,931	5,757
validation	4,822	2,106	2,003	713
test	4,714	2,066	1,982	666
total	48,414	21,362	19,916	7,136

origin	real 94.1% · synthetic 5.9%
platform	telegram 94.2% · facebook 2.2% · instagram 1.9% · youtube 1.7%
severity (0 / 1 / 2 / 3 / 4 / 5)	2.2% · 4.3% · 30.2% · 35.4% · 23.5% · 4.4%
toxicity types (multi-label)	profanity 63% · insult 59% · identity_attack 26% · implicit 21% · dehumanization 20% · threat 8% · sexual 8%
rewrite type	paraphrase 91.4% · light_edit 6.0% · brush_off 2.5%
median length in characters (original / neutral / literary)	93 / 97 / 121

Literary sub-styles

literary_style	modelled on	typical features	share
nechuy	Іван Нечуй-Левицький	calm, descriptive narrative prose, folk idiom	39.6%
franko	Іван Франко	argumentative, publicistic, measured irony	26.2%
lesya	Леся Українка	elevated, emotional, poetic	20.5%
kotlyarevsky	Іван Котляревський	burlesque, colloquial, playful	13.7%

The sub-styles are register targets, not imitations of specific texts.

Intended use

• Training and evaluating Ukrainian detoxification models (neutral or literary target).
• Studying text style transfer and register control in Ukrainian.
• severity and toxicity_types can support research on graded toxicity, but they are LLM annotations, not human labels.

Not intended for: generating toxic text, profiling or identifying commenters, or any commercial use.

Limitations

• The rewrites and the severity annotations are written and checked by LLMs. Audits were run on samples, and agreement between automatic judges and human review is moderate, so some rewrites lose nuance, soften the stance, or misread slang and names.
• The literary rewrites are the weakest part: some read as the neutral sentence with stock literary markers added, and a few markers («по щирості», «чоловіче добрий») are overused.
• The data is almost entirely Telegram comments on news and politics during the war, from 2022–2026. It over-represents war-related topics, political insults and profanity, and under-represents other platforms and topics.
• Synthetic comments are mostly mild to moderate; the most severe comments are all real.
• Neutral rewrites replace dehumanising nicknames for groups and people with neutral names. Whether a given replacement is appropriate is a judgement call.

Personal data

Comments are short public messages, but they were written by real people. User mentions, links and phone numbers are masked with [user], [url] and [phone]; channel names, author names, dates, message ids and the parent posts the comments replied to are not included. Comments may still name public figures or contain details that identify someone. If you find a row that identifies you or another private person, open a discussion on this repository and it will be removed.

Citation

@misc{uadetox2026,
  title  = {UADetox: Ukrainian toxic comments with neutral and literary rewrites},
  author = {Housto4},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Housto4/uadetox}},
  note   = {Project page: https://uadetox.pages.dev}
}