知识卡片

普通读书笔记卡 · 8671 · 生成模型:GPT-6

前缀适配训练层内虚拟上下文

理解等级:高级

Prefix tuning为多层注意力提供可学习的虚拟前缀状态,基础权重可保持冻结。它不同于仅添加可见自然语言指令,前缀也可能占注意力长度和缓存资源;具体空间占用依实现而定。

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

      As the figure shows, prefix parameters are added at each layer. Prefix-tuning is much more parameter-efficient than bottleneck adapters, taking up only 0.1% or less of a model’s parameters, as compared to adapters where it is usually 2% or more. However, prefix-tuning is harder to train effectively than adapters. Prefix-tuning also reduces the sequence length of the model in order to accommodate the virtual tokens.

    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:摘录中关于前缀适配训练层内虚拟上下文的定义、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:As the figure shows, prefix parameters are added at each layer. Prefix-tuning is much more parameter-efficient than bottleneck adapters, taking up only 0.1% or less of a model’s parameters, as compared to adapters where it is usually 2% or more. However, prefix-tuning is harder to train effectively than adapters. Prefix-tuning also reduces the sequence length of the model in order to accommodate the virtual tokens.