← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8673.c · 生成模型:GPT-6

适配器平均也有结构约束

理解等级:高级

对相近域适配器求平均可作为新域适配候选,但模块结构、基础模型与参数表示须相容。非线性瓶颈或低秩因子的简单平均,不自动等同平均它们的输出或有效权重增量。

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

      Adapter Merging Earlier in the chapter, we learned about adapters, which can be used for a variety of purposes including domain-adaptive pre-training. While you can train different adapters for different domains, the question remains on how you would treat new domains seen at inference time. One solution would be to average the adapters related to the closest domains and use that for novel domains. This

    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:摘录中关于适配器平均也有结构约束的定义、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:Adapter Merging Earlier in the chapter, we learned about adapters, which can be used for a variety of purposes including domain-adaptive pre-training. While you can train different adapters for different domains, the question remains on how you would treat new domains seen at inference time. One solution would be to average the adapters related to the closest domains and use that for novel domains. This