SeTox Dataset

SeTox is a Chinese neologism toxicity detection dataset for evaluating and training search-augmented safety models. It focuses on emerging terms, implicit expressions, and context-dependent online meanings.

Files

File	Records	Description
data/raw/neologism_dct.json	974	Release neologism lexicon.
data/train/train_tool.json	995	Tool-use SFT examples with search observations.
data/train/train_normal.json	600	Direct SFT examples.
data/eval/neologism_test.json	624	Main neologism toxicity test set.
data/eval/general_test.json	368	General safety test set.
results/summary/qwen2.5-7b_text2_neologism.json	2	7B checkpoint metric summary.
results/summary/qwen2.5-3b_text2_neologism.json	2	3B checkpoint metric summary.

The released neologism file preserves the user-provided 974-record source file, including seven duplicated source id values: 79, 132, 569, 591, 373, 427, and 730. We do not silently reindex the raw file; use list order when a stable per-record index is needed. Content fields are preserved from the cleaned source file.

Labels

Evaluation labels are binary:

• safe
• unsafe

The lexicon also includes fine-grained Chinese risk categories:

• 贬损攻击
• 成人内容
• 违法犯罪
• 违背道德
• 敏感话题

Data Quality Notes

The evaluation sets and 974-record lexicon pass the release validation checks, including the known duplicated source ids described above. The tool-use training file is the exact final training snapshot, but its search observation turns include search-result noise such as failed pages, translation or scraper pages, and advertising snippets.

Current heuristic audit on data/train/train_tool.json:

Group	Records	Rate
Any reviewed noise pattern	235 / 995	23.62%
Ad/contact-like noise	101 / 995	10.15%
Page-quality-only noise	134 / 995	13.47%

For details, see the code repository's docs/data_audit.md.

Limitations and Content Warning

This dataset is intended for research on Chinese neologism toxicity detection. It contains unsafe, toxic, adult, illegal, or otherwise sensitive expressions as part of the annotation and evaluation task. The tool-use training observations also contain noisy search snippets. Users should review downstream uses carefully and avoid treating the training observations as clean factual sources.

Usage

import json
from pathlib import Path

records = json.loads(Path("data/eval/neologism_test.json").read_text(encoding="utf-8"))
print(records[0].keys())

License

The released data package is distributed under the MIT License in the SeTox code repository.

Citation

Please cite the SeTox ACL paper:

• Paper: https://aclanthology.org/2026.acl-long.1602/
• DOI: https://doi.org/10.18653/v1/2026.acl-long.1602

@inproceedings{cui-etal-2026-setox,
  title = "{S}e{T}ox: Search-enhanced Reasoning for {LLM}-based Toxicity Detection over {C}hinese Internet Buzzwords",
  author = "Cui, Yiming and Zhang, Qinglin and Su, Xu and Min, Changyu and Hu, Shilin and Huang, Minlie",
  editor = "Che, Wanxiang and Nabende, Joyce and Shutova, Ekaterina and Pilehvar, Mohammad Taher",
  booktitle = "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
  month = jul,
  year = "2026",
  address = "San Diego, California, USA",
  publisher = "Association for Computational Linguistics",
  url = "https://aclanthology.org/2026.acl-long.1602/",
  doi = "10.18653/v1/2026.acl-long.1602",
  pages = "33018--33034"
}