Nemotron-Content-VISafe-v1

Dataset Description: VISafe is a Vietnamese-language AI safety evaluation probe dataset for testing model and guardrail behavior on safety-critical prompts. The current validated build contains 3,212 text probes in Vietnamese across jailbreak, toxicity, misinformation, prompt injection, Vietnam-specific political sensitivity, cybercrime, over-refusal, and privacy categories. The dataset combines translated probes from established English safety benchmarks with Vietnamese-native probes generated from deterministic templates and curated seed prompts. It is designed to evaluate whether models refuse harmful requests, avoid unsafe completion behavior, handle prompt-injection attempts, respond neutrally to sensitive political topics, and avoid over-refusing benign Vietnamese requests. LANGUAGE: Vietnamese (vi), native name: Tiếng Việt. ISO 639-3: vie. Script: Latin with Vietnamese diacritics. This dataset contains adversarial, harmful, sensitive, and policy-violating prompts intended for evaluation and red-teaming only. It should not be used to train general-purpose conversational models without strict safety alignment and governance review. This dataset is ready for both commercial and non-commercial use Dataset Owner(s): NVIDIA Corporation Dataset Creation Date: May 1, 2026 Version: Version 1.0 Previous Version(s): No previous version Relationship to Previous Version(s): N/A License/Terms of Use: NVIDIA SAMPLE DATA LICENSE Intended Usage: PRIMARY USE CASES: Vietnamese AI safety evaluation - Measuring model compliance with safety policies on Vietnamese prompts Guardrail and refusal testing - Evaluating whether moderation systems detect and block harmful Vietnamese requests Red-teaming and robustness testing - Stress-testing multilingual models against jailbreaks, prompt injection, obfuscation, and harmful instruction-following Vietnam-localized risk assessment - Covering Vietnamese cultural, political, social, cybercrime, privacy, and misinformation contexts Over-refusa