← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6509 · 生成模型:GPT-6

已有序列评分无需重新采样

理解等级:中级

对已有回答评分时,模型在每个位置读取固定前缀并给已知下一个 token 分配概率,不由采样器决定后续输入。它与生成时选择一个 token 并延续路径不同;应区分原始模型分布评分与经温度、过滤后的生成分布。

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

      As shown in figure 5.11, we compute the probability of each token given its preceding tokens. This is the same procedure as in figure 5.10, except that we repeat it for every token in the sequence. Unlike text generation, we are not producing new tokens here. Instead, we retrospectively look up the probability assigned to each token in the already generated sequence.

    参考来源

    - 位置:《Build a Reasoning Model (From Scratch)》第05章《自我修正与推理时扩展》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-5.html) - 结论依据:摘录中关于已有序列评分无需重新采样的定义、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:As shown in figure 5.11, we compute the probability of each token given its preceding tokens. This is the same procedure as in figure 5.10, except that we repeat it for every token in the sequence. Unlike text generation, we are not producing new tokens here. Instead, we retrospectively look up the probability assigned to each token in the already generated sequence.