← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6622.f · 生成模型:GPT-6

词表大小平衡压缩与学习信号

理解等级:高级

更多词表项可能把常见长片段压成较少token,但增加嵌入和输出层成本,也可能产生训练稀少的项。词表最优大小依语言、语料、模型和预算,不能只最大化压缩率。

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 vocabulary size, the fewer the number of tokens required to represent a given text, thus increasing the compression efficiency. Thus, for the same amount of training or inference compute, the language model can process more text. However, as the vocabulary size increases, there are more and more rare tokens with limited occurrences in the training data, and these rare tokens will have deficient representations.

    参考来源

    - 位置:《Designing Large Language Model Applications》第03章《词表与分词》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch03.html) - 结论依据:摘录中关于词表大小平衡压缩与学习信号的定义、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 vocabulary size, the fewer the number of tokens required to represent a given text, thus increasing the compression efficiency. Thus, for the same amount of training or inference compute, the language model can process more text. However, as the vocabulary size increases, there are more and more rare tokens with limited occurrences in the training data, and these rare tokens will have deficient representations.