知识卡片

普通读书笔记卡 · 6625.a · 生成模型:GPT-6

词片合并评分不是只看低频

理解等级:高级

常见教学WordPiece训练用相邻对频率除以两个单元频率乘积来排序,强调相对结合强度。单元低频不保证其对必定最高分,还要看共同出现次数;原始实现未完全公开时应标明这只是基于文献的说明。

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

      score = freq(a,b)/(freq(a) * freq(b)) This means that if a token pair is made up of tokens that individually have low frequency, they will be merged first. Figure 3-4 shows the merge priority and how the normalization by individual frequencies affects the order of merging. Figure 3-4. WordPiece tokenization During inference, merge rules are not used. Instead, for each pre-tokenized token in the input text,

    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:摘录中关于词片合并评分不是只看低频的定义、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:score = freq(a,b)/(freq(a) * freq(b)) This means that if a token pair is made up of tokens that individually have low frequency, they will be merged first. Figure 3-4 shows the merge priority and how the normalization by individual frequencies affects the order of merging. Figure 3-4. WordPiece tokenization During inference, merge rules are not used. Instead, for each pre-tokenized token in the input text,