RooLeX/Homework2-llm

Dataset Card for Russian Toxicity Detection Dataset

Dataset Details

Dataset Description

This dataset is designed for binary classification of toxic and non-toxic Russian texts. It was created as part of an educational project to build a real-time quality control system for customer support messages. The goal is to automatically distinguish messages containing rudeness, sarcasm, or aggression (toxic) from neutral, polite, or informative messages (non-toxic).

• Curated by: RooLeX
• Language(s): Russian (ru)
• License: MIT License

Dataset Sources

• Repository: https://huggingface.co/datasets/RooLeX/Homework2-llm

Uses

Direct Use

This dataset is intended for training and evaluating binary text classification models that detect toxicity in Russian-language texts. It can be used in content moderation systems, customer support quality control, social media monitoring, and similar applications.

Out-of-Scope Use

• The dataset is not suitable for detecting toxicity in languages other than Russian.
• It should not be used to make automated decisions about individuals without human oversight.
• The dataset is balanced (equal number of toxic and non-toxic examples), which may not reflect real-world class distributions. Models trained on it may require threshold tuning for production use.

Dataset Structure

Data Fields

• text (string): The text content of the message.
• label (int): Binary label indicating toxicity (0 = non-toxic, 1 = toxic).

Data Splits

The dataset is split into three subsets:

Split	Number of Examples
Train	59,451
Validation	7,431
Test	7,432

All splits are balanced (approximately equal number of toxic and non-toxic samples).

Dataset Creation

Curation Rationale

The dataset was created to address the problem of automatically detecting toxic messages in customer support communications. Toxic messages can cause significant reputational damage, and the goal is to build a fast, accurate model that can flag or block such messages before they are sent to clients.

Source Data

Toxic class (label = 1) was collected from publicly available datasets containing toxic Russian comments:

• AlexSham/Toxic_Russian_Comments — comments from the social network ok.ru
• NiGuLa/Russian_Inappropriate_Messages — inappropriate/rude messages
• textdetox/multilingual_toxicity_dataset — multilingual toxicity data (Russian subset)

Non-toxic class (label = 0) was sourced from various Russian-language datasets that are unlikely to contain toxic content:

• SetFit/amazon_massive_intent_ru-RU — voice assistant commands
• DeepPavlov/mtop_intent_ru — user queries and intents
• MTS-AI-SearchSkill/MTSBerquad — question-answer pairs (QA)
• mteb/ru_nlu_intent — intent classification texts

Data Collection and Processing

1. All source datasets were loaded using the Hugging Face datasets library.
2. For toxic sources, only examples labeled as toxic (or inappropriate) were selected.
3. For non-toxic sources, all examples were taken (with label 0).
4. The combined data was cleaned:
   • Empty strings and NaN values were removed.
   • Exact duplicate texts were dropped.
   • Texts with less than 3 words or more than 500 words were filtered out.
5. Classes were balanced using undersampling of the larger class.
6. The balanced dataset was split into train/validation/test sets in 80/10/10 proportion with stratification to preserve class balance.

Annotations

The dataset contains no additional annotations beyond the original labels from the source datasets. All labels are binary (toxic/non-toxic).

Personal and Sensitive Information

The dataset consists of text snippets from public sources and does not contain personally identifiable information (PII) such as names, addresses, phone numbers, etc. However, some toxic comments may contain offensive language.

Bias, Risks, and Limitations

Known Biases

• Domain bias: Toxic examples are predominantly from social media (ok.ru), while non-toxic examples come from voice assistants, QA systems, and intent data. This may cause the model to perform differently on texts from other domains (e.g., news, forums, email).
• Cultural bias: The data reflects the cultural and linguistic nuances of Russian-language internet communication, which may not generalize to other languages or regions.
• Labeling bias: The source datasets use different definitions of toxicity/inappropriateness, which may introduce inconsistencies.
• Balance bias: The dataset is artificially balanced, which does not reflect real-world toxicity prevalence.

Risks

• The model may produce false positives (flagging harmless messages as toxic) or false negatives (missing genuine toxicity), which can have business consequences.
• The model may be sensitive to sarcasm, irony, or context, which are hard to capture in short texts.

Recommendations

• Users should evaluate the model on domain-specific data before deployment.
• A human-in-the-loop review process is recommended for production use.
• The classification threshold should be tuned based on business priorities (e.g., lower threshold for stricter moderation).

Dataset Card Authors

RooLeX

Dataset Card Contact

• Hugging Face: RooLeX

Glossary

• Toxicity: Text that contains insults, threats, aggression, or rude language.
• Binary classification: A classification task with two possible outcomes (toxic vs. non-toxic).

Citation

BibTeX:

@misc{russian_toxicity_dataset,
  author = {RooLeX},
  title = {Russian Toxicity Detection Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/RooLeX/Homework2-llm}
}