README.md

metadata

license: cc-by-nc-sa-4.0
language:
  - ko
size_categories:
  - 10K<n<100K
pretty_name: 'Ko-ToxicWebQ: Korean Toxic Web Query Dataset'
dataset_info:
  features:
    - name: query
      dtype: string
    - name: sub_category
      dtype: string
    - name: manipulation
      dtype: int64
    - name: abbreviation
      dtype: int64
    - name: w/o_transformation
      dtype: int64
  splits:
    - name: train
      num_bytes: 2305606
      num_examples: 38175
    - name: test
      num_bytes: 575772
      num_examples: 9539
  download_size: 1067434
  dataset_size: 2881378
extra_gated_heading: ⚠️ Restricted Access Dataset
extra_gated_description: >-
  This dataset contains toxic and sensitive real-world web queries, including
  sexually explicit and legally restricted content. Automatic access approval is
  enabled.
extra_gated_prompt: >-
  By requesting access, you agree to use this dataset solely for research and to
  handle sensitive content responsibly.
extra_gated_button_content: Agree and Request Access
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*

Ko-ToxicWebQ: Korean Toxic Web Query Dataset

Ko-ToxicWebQ is a dataset consisting of 47,714 real-world Korean web search queries annotated for harmfulness and obfuscation strategies. The queries are extracted from Microsoft Bing search logs and labeled according to culturally specific toxicity categories and linguistic evasion tactics. The dataset is designed for evaluating moderation systems and training language models to handle adversarial or sensitive inputs in Korean.

⚠️ Warning: This dataset includes explicit, potentially offensive, and legally sensitive content. Researchers must take necessary precautions when using or sharing model outputs trained on this dataset.

Ethics Statement

We recognize that providing real examples of this content may be uncomfortable or distressing. These samples, however, are critical for developing robust detection systems that can mitigate harm and prevent the circulation of illicit content. We urge researchers to handle this dataset responsibly, with appropriate safeguards for user well-being.

Since conceptions of toxic or illegal content vary widely across jurisdictions, we emphasize that our annotations are grounded in South Korean laws and societal norms. We neither endorse nor condone the content found in these queries; our sole aim is to enable the development of more effective toxicity detection methods.

---

💡 Dataset Highlights

• Language: Korean (ko)
• Size: 47,714 queries
• Labels:
  • sub_category: One of 7 toxicity categories
  • manipulation, abbreviation, w/o_transformation: Binary flags for obfuscation strategies

---

🗂 Features

Name	Type	Description
query	string	Raw web search query in Korean
sub_category	string	One of {clean, gambling, sexuality, social, copyright, prostitution, others}
manipulation	int	1 if query contains character-level manipulation, else 0
abbreviation	int	1 if query uses abbreviation (e.g., ㅇㄷ → 야동), else 0
w/o_transformation	int	1 if query is plain without evasion, else 0

---

🔎 Category Description

Toxicity categories follow Korean regulatory standards (e.g., KCSC):

Category	Description
clean	Non-toxic query
gambling	Illegal betting or casino platforms
sexuality	Explicit adult content, erotic material
social	Sexual exploitation or adult streams on platforms
copyright	Torrenting, illegal streaming or software cracking
prostitution	Online sex solicitation or red-light area keywords
others	Hacking, self-harm, drug-related, or uncategorized

---

🔐 Gated Access (Automatic Approval)

This dataset is gated on Hugging Face due to the presence of harmful and sensitive content. However, automatic approval is enabled, and users only need to click to request access.

By accessing the dataset, you agree to:

• Use it solely for research or academic purposes
• Avoid public dissemination of raw toxic queries
• Apply safety filters if used in generation/inference tasks

---

📦 Usage Example

from datasets import load_dataset

dataset = load_dataset("TEAMREBOOTT-AI/Ko-ToxicWebQ", token="hf_your_token")
print(dataset["train"][0])

Term of Use

This dataset is partially based on the MS MARCO Web Search data and is provided strictly for non-commercial research purposes only. It is made available free of charge to promote research in artificial intelligence and related fields. You may only use this dataset if you agree to and comply with all of the following conditions:

Conditions of Use:

• This dataset must be used for non-commercial purposes only and may not be used, either directly or indirectly, in any commercial product or service.

• The datasets are provided "as is" without warranty and usage of the data has risks since we may not own the underlying rights in the documents. We are not be liable for any damages related to use of the dataset.