Tag
The paper introduces NSFT, a fine-grained parameter-efficient fine-tuning framework for MoE LLMs that refines adaptation from experts to sub-experts, demonstrating improved performance with fewer trainable parameters.