tremper01/ru-forum-toxicity-clean

Russian Forum Posts Clean

Dataset Description

This dataset contains cleaned Russian-language forum posts collected from public web pages.

The initial crawl produced 102,593 raw posts. After text cleaning, filtering, and deduplication, the dataset contains 83,075 unique cleaned posts.

This dataset is intended as a base corpus for experiments with Russian user-generated content, toxicity classification, weak supervision, text preprocessing, and lightweight text classification models.

This version does not contain toxicity labels. A separate weak-labeled version may be published later.

Dataset Details

• Language: Russian
• Domain: forum-style user-generated content
• Total rows: 83,075
• Format: Parquet
• Splits: train / validation / test
• Labels: none
• Text type: cleaned and deduplicated forum posts

Dataset Structure

The dataset contains the following columns:

Column	Type	Description
id	string	Unique row identifier
plane_text	string	Cleaned Russian forum post text
clean_text	string	Additional cleaned/normalized version of the text with service tokens where applicable
plane_text_length	int	Character length of text
clean_text_length	int	Character length of clean_text

Data Splits

Split	Rows
train	66,460
validation	8,307
test	8,308
Total	83,075

Data Files

text data/
├── train.parquet
├── validation.parquet
└── test.parquet 

Usage

python from datasets import load_dataset
dataset = load_dataset("tremper01/ru-forum-toxicity-clean")
print(dataset) print(dataset["train"][0]) 

Or with pandas:

python import pandas as pd
train = pd.read_parquet("data/train.parquet")
validation = pd.read_parquet("data/validation.parquet")
test = pd.read_parquet("data/test.parquet") 

Preprocessing

The raw collected posts were processed using a text cleaning pipeline.

The preprocessing included:

• removing empty or invalid posts;
• removing or normalizing noisy markup;
• normalizing whitespace;
• removing exact and near-exact duplicates after cleaning;
• preserving meaningful Russian text;
• preserving profanity and informal language where it is part of the original text.

The final dataset was deduplicated by cleaned plain text.

Intended Uses

This dataset can be used for:

• Russian text classification experiments;
• toxicity classification research;
• weak supervision pipelines;
• training lightweight language models or classifiers;
• studying noisy Russian user-generated content;
• evaluating preprocessing strategies for forum-style text.

Out-of-Scope Uses

This dataset should not be used for:

• harassment generation;
• user profiling;
• deanonymization;
• identifying or targeting individuals;
• fully automated moderation decisions without human review;
• high-stakes decision-making.

Limitations

• The dataset may contain offensive, toxic, explicit, or disturbing language.
• The dataset is based on forum-style user-generated content and may not generalize to all Russian-language domains.
• The dataset may contain cultural, stylistic, and topical biases from the source domain.
• Automatic cleaning may be imperfect.
• This version does not contain human-verified labels.
• This dataset should be treated as a research and educational resource, not as a production moderation dataset.

Ethical Considerations

The dataset consists of user-generated text collected from public web pages.

Direct identifiers and noisy technical artifacts were removed or normalized where detected. However, automatic filtering may be imperfect, and the dataset may still contain sensitive or offensive content.

The dataset is provided for research, educational, and experimental purposes. Users are responsible for ensuring that their use of the dataset complies with applicable laws, platform terms, and ethical guidelines.

Citation

If you use this dataset, please cite the dataset repository.

@dataset{
  ru_forum_posts_clean,
  title = {Russian Forum Posts Clean},
  author = {tremper01},
  year = {2026},
  publisher = {Hugging Face},
  note = {Cleaned Russian-language forum posts for text classification experiments}
}