This dataset contains offensive and abusive Hungarian online comments. It is released for non-commercial research use only. By requesting access you agree not to use it for targeted harassment, profiling of individuals, re-identification of pseudonymised users, or building text that evades moderation filters.

Hungarian Silver Toxicity Split

Warning: this dataset contains offensive, abusive and hateful language.

Dataset summary

148,474 Hungarian online comments in 35,431 threads, labelled for two pragmatic toxicity tests by an LLM ("silver" labels), split by thread into train/validation/test. It was released with Detecting Toxic Communication in Hungarian: Theory, Corpus, and Thread-Conditioned Classification (WOAH 2026) and used to fine-tune ModernBERT-HU-4k-thread-toxic.

The labels follow a three-test framework grounded in Culpeper's impoliteness theory:

T1, face-aggression: the comment uses a conventionalised or implicational impoliteness formula (insult, threat, condescension, silencer, curse) against the hearer's face, targeting the person rather than their opinion.
T3, community norm violation: the comment violates the behaviour that competent native participants expect in that platform and community.
T2 (antisocial perlocutionary effect, the Jigsaw "likely to make you leave" criterion) is not annotated here; it is covered by the earlier huBERT-SetFit classifier (Hatvani & Yang, 2025).

The two tests usually co-occur, but can come apart: friendly banter can be T1 without T3 (not toxic), and "toxic positivity" (warm wording that silences a real difficulty) is T3 without T1.

Source data

Comments come from five Hungarian platforms chosen to span registers and toxicity levels:

Source	Description
Reddit	39 Hungarian subreddits, collected with PRAW
YouTube	59 playlists centred on polarising public figures and news channels (YouTube Data API v3)
TikTok	Daily top Hungarian-tagged videos (TikTok Research API)
Kuruc.info	Hungary's most prominent radical outlet (Playwright scraping)
Mandiner.hu	Conservative news portal (Playwright scraping)

Threads were reconstructed from each platform's reply-to graph. Before annotation, personal names, usernames, e-mail addresses and URLs were replaced with typed placeholders (e.g. [person], [url]), keeping Hungarian case suffixes. Public figures were not pseudonymised.

Selection

From the assembled pool, only comments detected as Hungarian with confidence ≥ 0.5 by lingua-py and containing at least two Latin letters were eligible. Comments from the human gold set were excluded by exact match. Per source, items were drawn from a spread-sampled pool of up to 20,000 candidates until each of the four T1 × T3 label buckets held at least 500 annotations or the pool ran out. This is why Reddit, the largest source, contributes relatively few threads.

Annotation

All labels were produced by Gemini 3 Flash (google/gemini-3-flash-preview, temperature 0) with a Hungarian zero-shot prompt that defines T1 and T3 and asks for a JSON object with a boolean and a 1 to 4 confidence per test. The full prompt and its English translation are in the paper's appendix.

Gemini 3 Flash was chosen after benchmarking 11 LLMs against a 485-item human gold set annotated in full thread context. It had the best T1 F1 (0.769, κ = 0.648) and competitive T3 F1 (0.769, κ = 0.662). For comparison, human inter-annotator agreement on the gold set is α = 0.300 (T1) and α = 0.511 (T3). Like every evaluated LLM, it has high precision but lower recall on T1: it tends to label borderline implicational impoliteness as benign.

The secondary gold-annotation layers (target face type, impoliteness mechanism, pragmatic function, Hungarian-specific form) were not carried into the silver data because LLM-human agreement on them was too low.

Dataset structure

Splits

Split	Threads	Comments	% of comments
train	31,887	135,095	91.0%
validation	1,772	6,545	4.4%
test	1,772	6,834	4.6%

Data fields

Field	Type	Description
`comment_id`	string	Comment identifier. Unique only together with `thread_id` (see Known limitations).
`thread_id`	string	Conversation thread identifier, prefixed with the source (e.g. `kuruc:301109`).
`source`	string	Platform: `reddit`, `kuruc`, `mandiner`, `tiktok`, `youtube`.
`text_clean`	string	Cleaned, pseudonymised comment text (the annotated unit).
`toxic_weak`	int	Weak toxicity label from the huBERT-SetFit T2 classifier, used for pre-screening.
`toxic_score`	float	Score from the same weak classifier.
`t1_aggression`	bool	T1 face-aggression label.
`t1_confidence`	float (1-4)	LLM self-reported confidence for T1.
`t3_norm_violation`	bool	T3 community norm violation label.
`t3_confidence`	float (1-4)	LLM self-reported confidence for T3.
`parse_ok`	bool	Whether the LLM's JSON response parsed. True for every row.
`raw_response`	string	Raw LLM response.
`model`	string	Annotating model. Always `google/gemini-3-flash-preview`.

Label distribution

Label	yes	no
T1 (aggression)	37,632 (25.3%)	110,842 (74.7%)
T3 (norm violation)	40,808 (27.5%)	107,666 (72.5%)

Joint distribution (combo_cat, T1 / T3) by split:

Split	no / no	yes / yes	no / yes	yes / no	Total
train	97,488	33,972	3,262	373	135,095
validation	4,793	1,569	166	17	6,545
test	4,973	1,679	160	22	6,834
total	107,254 (72.2%)	37,220 (25.1%)	3,588 (2.4%)	412 (0.3%)	148,474

98.9% of T1-positive comments are also T3-positive. The mixed cells are small but present, as the framework predicts. The T1-only cell (0.3%) is rarer than in the human gold set, which suggests the LLM separates the two tests less finely than human annotators.

Suggested positive-class weights for training: T1 = 2.95, T3 = 2.64.

Confidence

Both confidence fields are skewed towards 4 (median 4, mean ≈ 3.9). Confidence is slightly lower for yes labels (mean ≈ 3.81) than for no labels (≈ 3.92 to 3.95).

Text length

Comment lengths in ModernBERT-HU-4k tokens: median 20, p90 82, p95 124, p99 250, p99.9 624, max 4,826. Fewer than 0.2% of comments exceed 512 tokens. Full threads are longer; the longest exceeds 2,000 tokens, which is why the paired model uses a 4,096-token context.

Split methodology

The split is at thread level, so no conversation is shared between partitions (verified: no thread_id overlap). Thread sizes are very skewed: 50.4% of threads have one comment, the largest has 672, and the top 1,000 threads hold 36% of rows. To keep the rare yes_no / no_yes combinations in every split, threads were stratified on whether they contain at least one such row, then split 90 / 5 / 5 with sklearn.model_selection.train_test_split(random_state=42). Every label combination stays within 1.5 percentage points of its global share in each split.

Known limitations

Silver labels only. One LLM produced every label; there is no within-dataset agreement signal. Use the human gold set for evaluation.
T1 under-recall. The annotating model is conservative on borderline face-aggression, and models trained on this data inherit that bias.
comment_id is not globally unique. For mandiner especially, IDs repeat across threads (e.g. man-comment-card_0 appears in 1,991 threads). Use (thread_id, comment_id) as the key.
No explicit turn order. There is no turn-index or parent field, so thread context must be reconstructed from row order within thread_id.
Thread concentration. A few very large threads (mostly kuruc and tiktok) account for a large share of rows.
No near-duplicate removal beyond the (thread_id, comment_id) key.

Intended use and license

Released for non-commercial research on Hungarian online abuse under CC BY-NC-SA 4.0, with gated access, primarily as a training resource for thread-aware Hungarian toxicity classifiers. It must not be used for targeted harassment, profiling of individuals, or building text that evades moderation filters.

The text is verbatim but pseudonymised; it is not paraphrased because the morphological markers under study do not survive paraphrase. If you find remaining personal data, please open a discussion on this repository.

Citation

Proceedings details (pages, anthology link) will be added once the WOAH 2026 proceedings are published.

@inproceedings{hatvani-madarasz-2026-detecting,
  title     = {Detecting Toxic Communication in {H}ungarian: Theory, Corpus, and Thread-Conditioned Classification},
  author    = {Hatvani, P{\'e}ter and Madar{\'a}sz, G{\'a}bor},
  booktitle = {Proceedings of the 10th Workshop on Online Abuse and Harms (WOAH 2026)},
  year      = {2026},
  address   = {Budapest, Hungary},
  publisher = {Association for Computational Linguistics}
}

Downloads last month: 39

Total file size:

21.5 MB

Models trained or fine-tuned on magyar-nlp-szine-java/silver-toxic-split

magyar-nlp-szine-java/ModernBERT-HU-4k-thread-toxic

Text Classification • 0.1B • Updated 18 days ago