知识卡片

普通读书笔记卡 · 8686.c · 生成模型:GPT-6

白盒蒸馏要明确表示对齐

理解等级:高级

拥有教师内部输出时,可匹配概率、隐藏状态或层级表示。概率分布可用明确方向的KL,普通激活需要合适距离或投影及维度对齐;不同层数和宽度不能仅凭同序号就视为语义相同。

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

      Intermediate representations This class of techniques is known as white-box distillation. Here the distillation dataset consists of intermediate representations from a model, which can include activations or output logits. This data can be used to align the student model with the teacher model. The alignment is learned using methods like KL-divergence, discussed in Chapter 4. Teacher feedback In this class of techniques, the outputs from

    参考来源

    - 位置:《Designing Large Language Model Applications》第09章《推理优化》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch09.html) - 结论依据:摘录中关于白盒蒸馏要明确表示对齐的定义、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:Intermediate representations This class of techniques is known as white-box distillation. Here the distillation dataset consists of intermediate representations from a model, which can include activations or output logits. This data can be used to align the student model with the teacher model. The alignment is learned using methods like KL-divergence, discussed in Chapter 4. Teacher feedback In this class of techniques, the outputs from