Dataset · nvidia
NVIDIA labeled dialog safety dataset
nvidia/Aegis-AI-Content-Safety-Dataset-1.0
Around 11,000 manually labeled dialogues for risk classification and LLM content moderation.
What's inside
An open dataset from NVIDIA (CC-BY-4.0 license) in English containing approximately 11,000 human-model dialogues. All interactions are manually labeled according to the NVIDIA safety taxonomy, which includes 13 critical risk categories.
What you can do with it
- Training safety classifiers and content moderation filters (guardrails) for LLMs
- Evaluating the vulnerability of language models to generating harmful content
- Tasks
- Text classification
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 1.7K
- Likes
- 63
Models trained on it
- nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0Text classification 1.8K
- nvidia/Aegis-AI-Content-Safety-LlamaGuard-Permissive-1.0Text classification 911
- blackmallangcow/nari-decision-4bEmbeddings 445
- KucLab/Barist.1Text classification 143
- KucLab/Censer.1Text classification 120
- KucLab/Barist.2Text classification 80
- KucLab/Censer.8Text classification 74
- KucLab/Censer.9Text classification 57
How to load it
from datasets import load_dataset
ds = load_dataset("nvidia/Aegis-AI-Content-Safety-Dataset-1.0")