知识卡片

普通读书笔记卡 · 8669.d · 生成模型:GPT-6

参数扩展需保护既有共享路径

理解等级:高级

增加层、专家或适配器可把部分新域学习放到新增参数。若仍更新共享注意力、嵌入或改变路由,旧行为依然可能变化;只有明确冻结与兼容路径等条件下,才能讨论旧分支的保持。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      Parameter Expansion An alternative to the replay approach is to use parameter expansion techniques. The naive way would be to just add a new layer or two on top of the model and train only those parameters during continual pre-training. You can also insert and train domain-specific parameter modules (called adapters) within existing layers. We will discuss adapter-based approaches in “Parameter-Efficient Fine-Tuning”. Leveraging DEMix Layers

    参考来源

    - 位置:《Designing Large Language Model Applications》第07章《进阶微调技术》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch07.html) - 结论依据:摘录中关于参数扩展需保护既有共享路径的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:Parameter Expansion An alternative to the replay approach is to use parameter expansion techniques. The naive way would be to just add a new layer or two on top of the model and train only those parameters during continual pre-training. You can also insert and train domain-specific parameter modules (called adapters) within existing layers. We will discuss adapter-based approaches in “Parameter-Efficient Fine-Tuning”. Leveraging DEMix Layers