Paper: WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Models
Data: DatasetHF_link
WildJailbreak is an open-source synthetic safety-training dataset with 262K vanilla (direct harmful requests) and adversarial (complex adversarial jailbreaks) prompt-response pairs. In order to mitigate exaggerated safety behaviors, WildJailbreaks provides two contrastive types of queries: 1) harmful queries (both… See the full description on the dataset page:
https://huggingface.co/datasets/walledai/WildJailbreak.
