Remove a specific person or conversation from an already-trained model on request — the consent theme made concrete.
Directions: unlearning methods (gradient ascent, SISA, adapter ablation); verify the data is gone via extraction tests. Part of the exploratory roadmap.
Remove a specific person or conversation from an already-trained model on request — the consent theme made concrete.
Directions: unlearning methods (gradient ascent, SISA, adapter ablation); verify the data is gone via extraction tests. Part of the exploratory roadmap.