Auxiliary-Loss-Free (arXiv:2408.15664) 详解:用 expert-wise bias + 规则式更新替代传统 auxiliary balance loss,消除干扰梯度对主任务训练的污染。balance 与 specialization 通过 bias 与 affinity 解耦。V3 训练全面采纳。
![]()
ESFT (Expert-Specialized Fine-Tuning, arXiv:2407.01906) 详解:基于 MoE 模型在下游任务上 expert 激活的天然稀疏性,只 fine-tune 任务相关的少数 expert,5-25% 可训练参数即可匹配 Full FT 性能,明显优于 LoRA。
![]()
转载本文请注明出处:https://yudonglee.me/deepseekm… Continue Reading →
![]()
© 2026 Yudong‘s Blog — Powered by WordPress
Theme by Anders Noren — Up ↑