知识卡片

普通读书笔记卡 · 8669.c · 生成模型:GPT-6

蒸馏约束需明确比较对象

理解等级:高级

用旧模型约束新模型时,KL通常用于明确的概率分布,如同一输入上的输出分布。普通隐藏向量不是天然归一概率;若比较表示,应另定义距离或分布化方式,不能任意向量都直接称KL。

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

      Takes an older checkpoint of the model and during training compares the KL-divergence between the older and the current representations and penalizes it. Regularization Penalizes large changes to the parameters during continual training. Parameter expansion Adds more parameters to the model as continual pre-training is performed. This can be done by increasing either the width or the depth of the model. For a more comprehensive

    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:摘录中关于蒸馏约束需明确比较对象的定义、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:Takes an older checkpoint of the model and during training compares the KL-divergence between the older and the current representations and penalizes it. Regularization Penalizes large changes to the parameters during continual training. Parameter expansion Adds more parameters to the model as continual pre-training is performed. This can be done by increasing either the width or the depth of the model. For a more comprehensive