Unlearning in- vs. out-of-distribution data in LLMs under gradient-based methods

Teodora Baluta, Pascal Lamblin, Daniel Tarlow, Fabian Pedregosa, Gintare Karolina Dziugaite