知识卡片

普通读书笔记卡 · 8664 · 生成模型:GPT-6

低秩适配的秩不是注意力维数

理解等级:高级

LoRA的r表示低秩增量的秩,alpha控制缩放,目标层决定哪些权重接受增量。它们不是原模型注意力头维度;冻结基础参数减少训练状态,但仍需基础模型前向与相应激活计算。

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

      The attention dimension of LoRA. lora_alpha The alpha parameter in the LoRA technique. lora_dropout The dropout probability used in the layers being tuned. This helps reduce overfitting. layers_to_transform This specifies the layers for which the LoRA transformation is to be applied. Here are some recommended default values: r = 64 lora_alpha = 8 lora_dropout = 0.1 For more background on LoRA, refer to Ogban Ugot’s

    参考来源

    - 位置:《Designing Large Language Model Applications》第06章《微调》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch06.html) - 结论依据:摘录中关于低秩适配的秩不是注意力维数的定义、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:The attention dimension of LoRA. lora_alpha The alpha parameter in the LoRA technique. lora_dropout The dropout probability used in the layers being tuned. This helps reduce overfitting. layers_to_transform This specifies the layers for which the LoRA transformation is to be applied. Here are some recommended default values: r = 64 lora_alpha = 8 lora_dropout = 0.1 For more background on LoRA, refer to Ogban Ugot’s