AliceYin/genz-contextual-abusive-slang-v2-silver

Gen Z Contextual Abusive Slang Benchmark v2 Silver

This is a research-only silver candidate dataset for studying Gen Z / internet-slang abusive-language classification. It is designed to test whether models can distinguish slang from abuse, profanity from harassment, meme mockery from benign meme use, and identity mentions from identity attacks.

This is not a final gold benchmark. Labels are model-assisted silver labels generated with a three-pass LLM annotation pipeline and adjudication. Use data/gold_candidate_balanced_2000.jsonl.gz as a candidate queue for human/expert adjudication before freezing a public leaderboard test set.

Content Warning

The dataset contains offensive language, profanity, harassment, identity-directed abuse, violent wishes, and sexualized abuse. It is intended for research on safety and language understanding. Do not use it to target, harass, profile, or deanonymize individuals or communities.

Dataset Files

Main benchmark split:

• data/train.jsonl.gz: 7,991 rows
• data/dev.jsonl.gz: 1,002 rows
• data/test.jsonl.gz: 1,007 rows

Quality tiers:

• data/silver_high.jsonl.gz: 5,045 rows
• data/silver_medium.jsonl.gz: 4,433 rows
• data/hard_review.jsonl.gz: 522 rows
• data/gold_candidate_balanced_2000.jsonl.gz: 2,000 rows

Full annotated set:

• data/full_with_benchmark_metadata.jsonl.gz: 10,000 rows

Supporting material:

• schema/annotation_schema_v2.json
• reports/llm_annotation_v2_expanded_fast_summary.json
• reports/v2_quality_split_report.json
• reports/baseline_v2_tfidf_logreg_report.json
• reports/baseline_v2_tfidf_logreg_silver_only_report.json
• docs/GENZ_LABELING_QUALITY_GUIDE.md
• docs/V2_RESULTS_REPORT_ZH.md
• scripts/build_v2_quality_splits.py
• scripts/train_v2_baseline.py

Annotation Schema

The Hugging Face release uses a flat schema so the Dataset Viewer and datasets library can infer stable features across heterogeneous source metadata. Complete nested objects are preserved as compact JSON strings in annotation_json, metadata_json, quality_json, benchmark_json, and llm_annotations_json.

Core flat fields:

• primary_label: one of clean, benign_genz_slang, profanity_non_abusive, non_targeted_offensive, targeted_insult_harassment, identity_attack_hate, threat_or_violent_wish, sexualized_abuse, ambiguous_context
• binary_abusive
• severity: integer from 0 to 4
• target_type: none, self, individual, public_figure, protected_group, non_protected_group, institution, object, unknown
• has_genz_slang
• slang_terms
• slang_function
• slang_freshness
• context_flags
• toxic_span
• slang_span
• confidence
• agreement
• needs_adjudication
• quality_split: silver_high, silver_medium, or hard_review
• source_bucket
• slices
• split
• annotation_json
• metadata_json
• benchmark_json
• llm_annotations_json

Label Distribution

Label	Count
clean	4,579
benign_genz_slang	2,183
targeted_insult_harassment	1,903
non_targeted_offensive	515
profanity_non_abusive	365
ambiguous_context	350
identity_attack_hate	44
threat_or_violent_wish	38
sexualized_abuse	23

Quality signals:

• agreement=1.00: 7,521
• agreement=0.67-0.99: 2,091
• agreement<0.67: 388
• needs_adjudication: 6
• has_genz_slang: 3,284
• has_slang_span: 3,698
• has_toxic_span: 3,133

Challenge Slices

Selected slice sizes:

Slice	Count
false_slang_hit	2,453
benign_slang	2,183
benign_profanity	365
targeted_insult	1,903
identity_hate	44
threat_or_violent_wish	38
sexualized_abuse	23
ambiguous_context	350
current_genz_slang	2,342
toxic_with_slang	884
benign_slang_contrast	2,400
aave_or_dialect	30
quoted_or_counterspeech	104
obfuscated_spelling	53

Baseline

TF-IDF word+character n-grams + Logistic Regression, test set:

Task	Accuracy	Macro-F1	Weighted-F1
primary_label 9-way	0.6763	0.3868	0.6844
binary_abusive	0.8381	0.7991	0.8423
target_type	0.7408	0.4147	0.7537
has_genz_slang	0.8908	0.8759	0.8905
severity	0.7249	0.3978	0.7279

Loading

from datasets import load_dataset

ds = load_dataset("YOUR_USERNAME/genz-contextual-abusive-slang-v2-silver", "benchmark")
print(ds)
print(ds["train"][0]["primary_label"])

Load the full annotated set:

full = load_dataset("YOUR_USERNAME/genz-contextual-abusive-slang-v2-silver", "full")

Load quality tiers:

quality = load_dataset("YOUR_USERNAME/genz-contextual-abusive-slang-v2-silver", "quality")

Intended Use

Appropriate uses:

• Research on abusive-language classification.
• Evaluating robustness to Gen Z slang, meme language, benign profanity, and false slang hits.
• Training bootstrap classifiers before human adjudication.
• Selecting candidates for a human gold benchmark.

Non-intended uses:

• Production moderation without independent validation.
• Enforcement actions against people or communities.
• Demographic profiling.
• Treating model-assisted labels as ground truth.
• Claiming leaderboard SOTA on the silver labels as if they were gold human labels.

Known Limitations

• This is a silver dataset, not a final gold benchmark.
• Rare positive classes are thin: identity hate, threats, and sexualized abuse need targeted expansion.
• Source shortcuts exist. Some source buckets are mostly benign while others are toxicity-heavy.
• AAVE/dialect and reclaimed/in-group slices are too small for strong fairness claims.
• LLM-assisted labels can encode model bias and should be audited before public benchmark use.

Data Use and License

License is marked as other because source terms vary across the underlying public/derived text sources. Use this release for research and comply with the terms of the original data sources and platforms. Do not redistribute in a way that violates source dataset or platform terms.

Citation

No formal citation is available yet. If this dataset becomes public, add a dataset card citation and paper/preprint reference before announcing it.