Dataset · allenai
Examples for safe LLM training and moderation
allenai/wildguardmix
Text collection annotated with harmful and safe prompts and responses for LLM safety research.
What's inside
WildGuardMix contains examples that may be disturbing, harmful, or upsetting: discrimination, violence, self-harm, sexual content, misinformation, and other risky categories. The goal is to advance research on safe LLM creation. English data, size 10K to 100K examples; modalities include tables and text. Training LLMs solely on harmful examples is not recommended.
What you can do with it
- Train prompt and response safety classifiers
- Evaluate LLMs or moderators on harmful content
- Fine-tune LoRA for safety filtering
- Tasks
- Text classification
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 11.3K
- Likes
- 130
- Access
- on request on Hugging Face
Models trained on it
- allenai/wildguardText generation 177.7K
- mradermacher/YuFeng-XGuard-Reason-8B-GGUF 1.2K
- Alibaba-AAIG/YuFeng-XGuard-Reason-8BText generation 779
- Alibaba-AAIG/YuFeng-XGuard-Reason-0.6BText generation 774
- mradermacher/Spellguard-1.0-8B-GGUF 754
- mradermacher/Wildguard-Qwen3-4b-GGUF 682
- mradermacher/YuFeng-XGuard-Reason-0.6B-GGUF 452
- mradermacher/GPT-Distill-Qwen3-8B-Spellguard-SLERP-GGUF 337
How to load it
from datasets import load_dataset
ds = load_dataset("allenai/wildguardmix")