Evil Sharpens Good.
Developed By: The LAION-AI Safety Team
Blade2Blade is a system that preformes fully automated redteaming on any chat, or classifcation model. By using RL, we tune an adversarial user prompter to attack other models and create user prompts that promote dangerous responses from the attacked model.
Below shows an example of a Blade2Blade type system that attacks Good T5, a model that is designed to predict Rules of Thumb and safety labels from the prosocial dialogue dataset (https://huggingface.co/datasets/allenai/prosocial-dialog).
