Ukrainian Tweets for Binary Toxicity Classification, 5k Crowdsourced Examples

5,000 Ukrainian tweets were filtered and labeled via the Toloka.ai crowdsourcing platform, resulting in a balanced set of 2.5k toxic and 2.5k non-toxic texts. This dataset is part of the larger multilingual_toxicity_dataset collection. It was uploaded by ukr-detect and last updated in November 2024.

Use Cases

• Train binary classifiers to detect toxic content based on the labeled Ukrainian tweets.
• Benchmark model performance on a balanced, crowdsourced toxicity dataset.
• Fine-tune multilingual language models for Ukrainian-specific moderation tasks.
• Study linguistic patterns of toxicity in Ukrainian social media text.

Strengths

• Contains a balanced set of 2,500 toxic and 2,500 non-toxic examples.
• Labels were created via a structured crowdsourcing task on the Toloka.ai platform.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is limited to 5,000 examples, which may be insufficient for some deep learning approaches.
• Data may reflect temporal and platform bias inherent to its source of Twitter posts.

Provenance

Source

ukr-detect on Hugging Face

Collection Method

Filtered from a Ukrainian Tweets dataset and labeled via Toloka.ai crowdsourcing.

Freshness

Last updated 2024-11-08 12:42:27

Geography

Ukraine (implied by language)