Train so the model provably memorises less — directly attacking the central leakage risk of training on personal chats.
Directions: DP-SGD / DP fine-tuning; privacy/utility trade-off vs. the style eval; combine with redaction. Part of the exploratory roadmap.
Train so the model provably memorises less — directly attacking the central leakage risk of training on personal chats.
Directions: DP-SGD / DP fine-tuning; privacy/utility trade-off vs. the style eval; combine with redaction. Part of the exploratory roadmap.