Simple data augmentation reduces blackmail attempts in models

In a developers' test, they simply augmented a chat training dataset aimed at harm reduction with independent tools and system messages; this more quickly reduced the rate of blackmail responses. The result shows that data diversification is an effective tool for improving model security.