知识卡片

普通读书笔记卡 · 6746.b · 生成模型:GPT-6

字节覆盖不等于通用语义覆盖

理解等级:中级

完整256字节基底可表示UTF-8编码,不必为每个Unicode字符单独建词表。新字符可能分成更多token,数量增幅依文本而变;不能由此断言固定多70%或能理解所有语言,字符数量也不同于语言数量。

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

      #3 A ByteLevel tokenizer is less efficient but more robust (no OOV tokens) than a character-level tokenizer. The ByteLevel part of your BPE tokenizer ensures your tokenizer will never miss a beat (or byte) as it is tokenizing your text. A byte-level BPE tokenizer can always construct any character by combining one of the 256 possible single-byte tokens available in its vocabulary. This means it

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第09章《可堆叠的Transformer》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-9.html) - 结论依据:摘录中关于字节覆盖不等于通用语义覆盖的定义、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:#3 A ByteLevel tokenizer is less efficient but more robust (no OOV tokens) than a character-level tokenizer. The ByteLevel part of your BPE tokenizer ensures your tokenizer will never miss a beat (or byte) as it is tokenizing your text. A byte-level BPE tokenizer can always construct any character by combining one of the 256 possible single-byte tokens available in its vocabulary. This means it