magyar-nlp-szine-java/silver-toxic-split

Hungarian Silver Toxicity Split

Dataset summary

This dataset contains 148,474 Hungarian online comments annotated for two pragmatic-relational toxicity dimensions — T1 (aggression) and T3 (norm violation) — using LLM-generated ("silver") labels. The dataset is pre-split into train/validation/test partitions, designed for fine-tuning hun-modernBERT-base_4k on multilabel aggression / norm-violation classification.

Unlike a small, human-annotated gold dataset covering the same dimensions, this silver dataset is annotated entirely by a single LLM (google/gemini-3-flash-preview) and is intended primarily as a larger-scale training resource rather than an evaluation benchmark.

Dataset structure

Splits

Split	Rows	% of total
train	135,095	91.0%
validation	6,545	4.4%
test	6,834	4.6%

Data fields

Field	Type	Description
comment_id	string	Comment identifier. Not guaranteed globally unique on its own (see Known limitations); unique only in combination with thread_id.
thread_id	string	Identifier of the conversation thread the comment belongs to.
source	string	Origin platform of the comment (reddit, kuruc, mandiner, tiktok, youtube).
text_clean	string	Cleaned comment text (the unit that was annotated).
toxic_weak	int/bool	Pre-existing weak toxicity signal (not produced by the T1/T3 annotation pipeline).
toxic_score	float	Pre-existing weak toxicity score (not produced by the T1/T3 annotation pipeline).
t1_aggression	bool	Whether the comment contains a conventionalized aggressive form (insult, threat, condescension, curse, personal attack).
t1_confidence	int (1–4)	LLM self-reported confidence for the t1_aggression label.
t3_norm_violation	bool	Whether the comment violates the expected community/contextual norms.
t3_confidence	int (1–4)	LLM self-reported confidence for the t3_norm_violation label.
parse_ok	bool	Whether the LLM's JSON response was successfully parsed. All rows in this dataset have parse_ok=True.
raw_response	string	Raw LLM API response (including reasoning trace, where available).
model	string	Annotating model identifier. All rows: google/gemini-3-flash-preview.

Label distribution

Computed over the full dataset (148,474 rows, all with valid t1_aggression / t3_norm_violation labels):

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 (T1, T3) distribution:

Combination	Count	%
no / no	107,254	72.2%
yes / yes	37,220	25.1%
no / yes	3,588	2.4%
yes / no	412	0.3%

Note: T1 and T3 are highly correlated in this silver-labeled dataset (agreement on ~97.3% of rows). The yes_no combination — aggressive but not norm-violating — is notably rare (0.3%) compared to a smaller human-annotated gold dataset covering the same phenomena, where this combination is substantially more common. This suggests the annotating LLM may distinguish the two dimensions less finely than human annotators; this is worth accounting for when interpreting model performance trained on this data, or when comparing against gold-standard evaluation.

Confidence scores

t1_confidence and t3_confidence use a 1–4 scale and are strongly skewed toward 4 (median = 4 for both, mean ≈ 3.9). Confidence is slightly lower on average for yes labels (mean ≈ 3.81) than for no labels (mean ≈ 3.92–3.95), consistent with positive/harder cases being less certain for the model.

Text length

Token lengths were computed using the hun-modernBERT-base_4k tokenizer (max length 8,192).

Statistic	Value
Median	20 tokens
p90	82 tokens
p95	124 tokens
p99	250 tokens
p99.9	624 tokens
Max	4,826 tokens

Fewer than 0.2% of rows exceed 512 tokens, and none exceed the model's 8,192-token limit. The dataset's effective sequence length is short relative to the model's context window — most rows are single, comment-level utterances rather than long threads.

Split methodology

The dataset was split at the thread level (thread_id), not the row level, to prevent context leakage: comments from the same thread can include preceding-turn context, so splitting individual comments across train/validation/test could let a model see fragments of the same conversation in multiple partitions.

Thread sizes are highly skewed: 50.4% of threads contain a single comment, while the largest threads (mostly from kuruc and tiktok) contain several hundred comments each (max: 672). The top 1,000 threads alone account for 36% of all rows.

To preserve representation of the rare yes_no / no_yes label combinations across all three splits, threads were stratified on a binary indicator (has_rare: whether the thread contains at least one row with combo_cat in {yes_no, no_yes}) before a grouped train/validation/test split (90% / 5% / 5%) was performed with sklearn.model_selection.train_test_split, random_state=42.

Validation of the resulting split (row-level, post-split):

Split	no_no	yes_yes	no_yes	yes_no	Max deviation from global ratio
train	72.16%	25.15%	2.41%	0.28%	0.08 pp
validation	73.23%	23.97%	2.54%	0.26%	1.10 pp
test	72.77%	24.57%	2.34%	0.32%	0.53 pp

No thread_id overlap exists between any pair of splits. source distribution is broadly similar across splits (e.g. kuruc: 30.5–33.8%, reddit: 19.8–24.1%), with minor variation attributable to the thread-level grouping of a few very large threads.

Known limitations

• Silver labels only: all T1/T3 labels were produced by a single LLM (google/gemini-3-flash-preview), not human annotators. No inter-annotator agreement signal is available within this dataset.
• comment_id is not globally unique: for the mandiner source in particular, comment_id values follow a template/positional convention that repeats across threads (e.g. man-comment-card_0 appears under 1,991 different thread_ids). The only globally unique key is the (thread_id, comment_id) pair.
• T1/T3 correlation: as noted above, T1 and T3 are very highly correlated in this dataset (Jaccard-like joint agreement ~97%), more so than in a comparably-scoped human-annotated gold dataset. Models trained primarily on this data may underfit the distinction between "aggressive but not norm-violating" language.
• Source imbalance and thread concentration: a small number of very large threads (mostly kuruc and tiktok) account for a disproportionate share of rows, which may bias source-level representation despite the thread-stratified split.
• No deduplication of near-identical text was performed beyond the (thread_id, comment_id) key check.

Intended use

Primary intended use is supervised fine-tuning of Hungarian encoder models (e.g. hun-modernBERT-base_4k) for multilabel aggression / norm-violation classification. Given the label correlation noted above, this dataset is best used as a large-scale training resource, complemented by a smaller human-annotated gold dataset for evaluation and for capturing the finer-grained distinction between T1 and T3.