知识卡片

普通读书笔记卡 · 6622.e · 生成模型:GPT-6

字符编码用更长序列换覆盖

理解等级:中级

细到字符或字节可降低某些词表外问题,却常使同一文本序列更长,增加窗口占用与计算。具体成本还受架构和下采样设计影响,不能只凭词表变小就断言整体模型更省资源。

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

      number of tokens needed to represent the average sentence becomes much larger. For example, the sentence, “The number of tokens needed to represent the average sentence becomes much larger,” contains 13 tokens when you treat each word as a token, but 81 tokens when you treat each character as a token. This reduces the amount of content you can represent within a fixed sequence length,

    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:摘录中关于字符编码用更长序列换覆盖的定义、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:number of tokens needed to represent the average sentence becomes much larger. For example, the sentence, “The number of tokens needed to represent the average sentence becomes much larger,” contains 13 tokens when you treat each word as a token, but 81 tokens when you treat each character as a token. This reduces the amount of content you can represent within a fixed sequence length,