Guardrail-TR: Turkish Prompt Safety Dataset with Hazard Labels

Turkish prompt-safety dataset containing approximately 405,000 single-turn user prompts. Each prompt has a binary safe/unsafe label and multi-label hazard categories, with all text in Turkish. The dataset was created by ytu-ce-cosmos and last updated on July 20, 2026.

Use Cases

• Train safety classifiers based on binary safe/unsafe labels.
• Evaluate multi-label hazard classification models based on hazard categories.
• Benchmark Turkish-language guardrail systems based on the large-scale prompt collection.
• Study cross-cultural content moderation based on translated and culturally adapted prompts.

Strengths

• Approximately 405,000 single-turn prompts provide a substantial corpus.
• Includes both binary safety labels and multi-label hazard categories for granular analysis.
• All prompt text is in Turkish, focusing on a specific language domain.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect source or translation bias inherent to the collection method.

Provenance

Source

ytu-ce-cosmos

Collection Method

Likely contains prompts processed and translated from English sources, with cultural adaptation.

Freshness

Last updated 2026-07-20 10:47:59; freshness should be verified.