Kinopotok Toxicity: Russian Customer Support Messages for Toxicity Classification

A dataset for binary classification of toxicity in user messages to an online cinema's support chat. Toxic messages were synthetically generated by a local Qwen/Qwen3-4B-Instruct-2507 model across six categories, while non-toxic messages were extracted from MTEB BANKING77 and two Bitext datasets, filtered, and translated into Russian. The dataset was created by aurelianvolturi and last updated on July 12, 2026.

Use Cases

• Train a binary classifier to detect toxic messages in Russian support chats.
• Benchmark model performance on synthetically generated toxic text across six categories.
• Fine-tune language models for content moderation in Russian-language online services.

Strengths

• Designed for a specific, practical application: toxicity detection in customer support.
• Toxic examples were systematically generated across six distinct categories.
• Non-toxic examples were sourced from established datasets and translated into Russian.

Limitations

• Row count is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• The synthetic generation of toxic messages may not fully capture real-world linguistic patterns.

Provenance

Source

huggingface

Collection Method

Toxic messages synthetically generated by a Qwen model; non-toxic messages sourced from MTEB BANKING77 and Bitext datasets, filtered and translated.

Time Range

null

Freshness

Last updated 2026-07-12 19:38:30; freshness should be verified.

Geography

null