知识卡片

普通读书笔记卡 · 8702 · 生成模型:GPT-6

套娃训练让前缀维度可用

理解等级:高级

MRL在多个前缀维度分别计算训练损失并加权,让短前缀也具备目标任务表示能力。使用时可按训练支持的维度裁剪并按要求归一;任意普通嵌入截前128维不具备同样训练保证。

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

      Matryoshka Representation Learning (MRL). In MRL, we first choose a set of truncation dimensions. For example a 1,024-dimension vector can have truncation dimensions 128, 256, 512, and 768. During the training process, we calculate the loss over each of the truncation dimensions as well as the full dimension. The losses are then added and weighted. In our example, the first 128 dimensions learn from the

    参考来源

    - 位置:《Designing Large Language Model Applications》第11章《表示学习与嵌入》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch11.html) - 结论依据:摘录中关于套娃训练让前缀维度可用的定义、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:Matryoshka Representation Learning (MRL). In MRL, we first choose a set of truncation dimensions. For example a 1,024-dimension vector can have truncation dimensions 128, 256, 512, and 768. During the training process, we calculate the loss over each of the truncation dimensions as well as the full dimension. The losses are then added and weighted. In our example, the first 128 dimensions learn from the