Catastrophic interference

New method recovers lost language model capabilities without retraining

Researchers have developed a novel post-hoc method called DG-Hard to address catastrophic forgetting in language models. This technique aims to recover lost capabilities after fine-tuning without requiring retraining, b…