Dataset · allenai
Synthetic dataset for conversational model safety
allenai/wildjailbreak
262K synthetic prompt-response pairs: harmful and similar harmless prompts for safety training.
What's inside
WildJailbreak is an open synthetic safety dataset containing 262K prompt-response pairs in vanilla and adversarial variants. To reduce over-refusals, it includes two contrastive prompt types: harmful and harmless, similar in form but without harmful intent. Language is English; modalities in metadata are text and images.
What you can do with it
- SFT/RLHF fine-tuning of safe assistants
- Adjust refusal and helpfulness balance
- Red teaming and jailbreak evaluation
- Tasks
- Text generation
- Languages
- English
- Size
- 1K–10K rows
- Downloads this month
- 7.5K
- Likes
- 175
- Access
- on request on Hugging Face
Models trained on it
- patronus-studio/wolf-defender-prompt-injectionText classification 10K
- patronus-studio/wolf-defender-prompt-injection-smallText classification 3.8K
- mradermacher/Spellguard-1.0-8B-GGUF 754
- mradermacher/GPT-Distill-Qwen3-8B-Spellguard-SLERP-GGUF 337
- AxiomicLabs/X3Guard-150MText classification 294
- patronus-studio/wolf-defender-prompt-injection-edgeText classification 215
- vllm-sr/Decision-1.0-Route-0.6BZero shot classification 101
- Spellguard/Spellguard-1.0-8BText generation 81
How to load it
from datasets import load_dataset
ds = load_dataset("allenai/wildjailbreak")