PolygloToxicityPrompts is a multilingual toxicity evaluation benchmark curated from web text. The dataset includes three splits: ptp-full, ptp-small, and wildchat, containing 25,000, 5,000, and 1,000 prompts per language, respectively. It was created by ToxicityPrompts and last updated on 2026-05-29.
Use Cases
- Benchmarking toxicity detection models based on multilingual prompts.
- Evaluating the safety of large language models across different languages.
- Training content moderation filters using web-sourced text examples.
- Comparing model performance across different dataset splits (full, small, wildchat).
Strengths
- Offers three distinct splits with defined sizes: 25K, 5K, and 1K prompts per language.
- Includes a split derived from AI2's WildChat dataset, providing a source of AI-generated conversation data.
- Multilingual scope supports evaluation across numerous languages.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count per split is known, but total dataset size and file formats are unspecified.
- Freshness should be verified as the last update timestamp is from 2026.
Provenance
- Source
- ToxicityPrompts
- Collection Method
- Curated from web text; the wildchat split is created using AI2's WildChat dataset.
- Freshness
- Last updated 2026-05-29 03:59:10.