❗️Warning: this document contains content that may be offensive or upsetting.

KOTOX, the first Korean dataset for deobfuscation and detoxification, was built using linguistically grounded transformation rules to handle obfuscated toxic expressions.

KOTOX is the first Korean dataset designed for deobfuscation and detoxification of toxic language. Built from linguistically grounded transformation rules, it introduces obfuscated instances to model how users disguise offensive expressions in Korean.

Unlike previous datasets that focus mainly on English or clean toxic text, KOTOX captures diverse obfuscation patterns specific to Hangul and Korean phonology, such as phonological, iconological, transliteration-based, syntactic, and pragmatic variations.

It supports three complementary tasks—obfuscated toxic text classification, neutral text deobfuscation, and toxic text sanitization—providing a unified benchmark to evaluate the robustness of language models against obfuscated toxic content and to advance safer language technologies for low-resource languages.

The dataset enables three complementary tasks:

1. 🧠 Obfuscated Toxic Text Classification

• Classify whether an obfuscated sentence is toxic or neutral.

2. 🔤 Neutral Text Deobfuscation

• Restore an obfuscated neutral sentence to its original, clean form.

3. 🧼 Obfuscated Toxic Text Sanitization

• Rewrite obfuscated toxic text into a deobfuscated, neutral sentence while preserving meaning.

KOTOX defines 17 transformation rules across 5 linguistic approaches, based on Korean linguistic properties.

• Started from K/DA - 7.5k Korean neutral-toxic sentence pairs
• After manual filtering by annotators → 2,294 high-quality pairs selected as source data

• Applied transformation rules to both neutral and toxic sides of each pair
• Used an alogorithm to sample and apply 2-4 rules per text, depending on difficulty
  Easy: 2 rules
  Normal: 3 rules
  Hard: 4 rules

• Final dataset: 6.9k neutral-toxic-pairs + corresponding obfuscated counterparts
• Split into train/validation/test = 8:1:1 ratio for each difficulty level

└── data
    ├── KOTOX
    │   ├── easy
    │   ├── normal
    │   ├── hard
    │   └── total
    └── KOTOX_classification
        ├── easy
        ├── normal
        ├── hard
        └── total

KOTOX: for obfuscation and detoxification, 🤗 huggingface-KOTOX

KOTOX_classification: for toxic hate speech detection, 🤗 huggingface-KOTOX-classification

Install the necessary dependencises using the provided requirements

$ pip install -r requirements.txt

Add .env file for using OpenAI API

OPENAI_API_KEY= <Your OpenAPI Key>

Git colne G2P

$ git clone https://github.com/seongmin-mun/KoG2Padvanced.git

$ python augmentation.py

Modify the classification/train_config.py file.

$ cd classification
$ python train.py

Modify the classification/eval_config.py file.

$ cd classification
$ python eval.py

Modify the finetuning/train_examples.sh file.

$ chmod+x finetuning/train_examples.sh
$ ./train_examples.sh

@misc{lee2025kotoxkoreantoxicdataset,
      title={KOTOX: A Korean Toxic Dataset for Deobfuscation and Detoxification}, 
      author={Yejin Lee and Su-Hyeon Kim and Hyundong Jin and Dayoung Kim and Yeonsoo Kim and Yo-Sub Han},
      year={2025},
      eprint={2510.10961},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.10961}, 
}