Knowledge That Is Linear Before Fine-Tuning Generalizes After It
This study distinguishes active facts that support generalization from inert facts during language model fine-tuning.
This study distinguishes active facts that support generalization from inert facts during language model fine-tuning.