alexomur/toxic-texts

---
language:
- ru
task_categories:
- text-classification
pretty_name: toxic-texts
size_categories:
- 100K<n<1M
configs:
- config_name: default
data_files:
- split: train
path: data/train-00000-of-00001.parquet
---

# toxic-texts

## Description

An anonymized Russian-language toxicity classification dataset.

The dataset contains short and medium-length user-generated texts labeled for binary toxicity classification.
It was assembled from multiple upstream sources and then manually reviewed and partially relabeled.
For the Habr-derived subset, raw texts were taken from the upstream source and labeled manually for this dataset.

## Dataset structure

Columns:

- `id`: internal record id in this published dataset
- `text`: anonymized text content
- `label`: binary toxicity label, where `0` means non-toxic and `1` means toxic
- `source`: source domain bucket used during curation (`dvach`, `habr`, `ok`)
- `text_length`: character length of the anonymized text

## Anonymization

The published version removes or masks common direct identifiers and references, including:

- email addresses
- URLs
- domain mentions inside text
- `@mentions`
- phone-like strings

The `source` field is preserved as a coarse provenance signal.
Direct upstream record identifiers and source URLs are not included in the published dataset.

## Curation notes

This dataset is a curated derivative work.

- Part of the corpus was based on `marriamaslova/toxic_dvach_small` and manually relabeled in places.
- Part of the corpus was built using source material associated with `khvatov/ru_toxicity_detector` research lineage and then manually reviewed within this project.
- Raw texts from `IlyaGusev/habr` were used as source material for a separately performed manual annotation pass.

## Upstream references

- [marriamaslova/toxic_dvach_small](https://huggingface.co/datasets/marriamaslova/toxic_dvach_small)
- [khvatov/ru_toxicity_detector](https://huggingface.co/khvatov/ru_toxicity_detector)
- [IlyaGusev/habr](https://huggingface.co/datasets/IlyaGusev/habr)

## Intended use

This dataset is intended for research, experimentation, and model evaluation for Russian toxicity detection.
Users should independently review suitability, bias, and legal constraints before production or redistribution use.

## Limitations

- Labels are binary and reflect this project's working definition of toxicity.
- Source distributions are imbalanced.
- Text anonymization reduces direct identifiers but does not guarantee perfect privacy protection in every edge case.
- The dataset may retain social, linguistic, and platform-specific biases present in the source material.