Dataset · nvidia
NVIDIA multilingual response safety dataset
nvidia/Nemotron-Safety-Guard-Dataset-v3
Over half a million examples in 12 languages for training LLM safety filters (guard models).
What's inside
The dataset contains about 514,617 text examples in 12 languages, including English, German, French, Spanish, Arabic, Chinese, and Hindi. The data is predominantly generated synthetically using the CultureGuard pipeline for content safety evaluation.
What you can do with it
- Training multilingual moderator models to filter LLM queries and responses
- Classifying unsafe content in 12 languages
- Evaluating language model robustness against toxic or dangerous topics
- Tasks
- Text classification, Text generation
- Languages
- English, Arabic, German, Spanish, French, Hindi, Italian, Japanese and 4 more
- Size
- 100K–1M rows
- Downloads this month
- 4K
- Likes
- 38
Models trained on it
- nvidia/Llama-3.1-Nemotron-Safety-Guard-8B-v3Text generation 8.5K
- vllm-sr/Vela-2.0-0.8BZero shot classification 2K
- vllm-sr/Vela-2.0-0.3BZero shot classification 2K
- Horizon-Labs/content-safety-guard-smallText classification 1.1K
- vllm-sr/Vela-1.0-Encoder-307M-ShieldText classification 762
- mradermacher/Llama-3.1-Nemotron-Safety-Guard-8B-v3-GGUF 498
- Horizon-Labs/content-safety-guard-baseText classification 341
- vllm-sr/Vela-2.0-9BZero shot classification 248
How to load it
from datasets import load_dataset
ds = load_dataset("nvidia/Nemotron-Safety-Guard-Dataset-v3")