知识卡片
完整字节基底解决表示覆盖
字节级BPE可用完整256字节基底表示合法文本的编码,避免因未见字符缺少基底符号。一个Unicode字符可能由多个字节token组成,单独解码某个token未必得到完整字符;支持覆盖也不保证模型理解所有语言。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 词表与分词
原书出处 · 1 处
- · 查看原书依据
256 tokens, representing all the characters that can be represented by a byte. This ensures that every Unicode character can be encoded just by the concatenation of the constituent byte tokens. Hence, it also ensures that we will never encounter an <UNK> token. The GPT family of models use this tokenizer.
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:摘录中关于完整字节基底解决表示覆盖的定义、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:256 tokens, representing all the characters that can be represented by a byte. This ensures that every Unicode character can be encoded just by the concatenation of the constituent byte tokens. Hence, it also ensures that we will never encounter an <UNK> token. The GPT family of models use this tokenizer.