知识卡片

普通读书笔记卡 · 6762.d · 生成模型:GPT-6

束搜索保留有限候选而非全局保证

理解等级:中级

每步扩展保留若干候选前缀,再按累计分数与长度等规则筛选。束宽有限会剪掉潜在最优路径,因此不保证找到全空间最高概率句,也不保证概率高就事实更好;宽度、长度和成本需共同评价。

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

      Fortunately, greedy search and repetition filtering are not the only approaches you can take to generate the next token. Two ways to make the token decoding both less predictable and less noisy (incorrect) are beam search and sampling. Beam search keeps track of several possible “beams” or paths through the fishbone diagram. That way, you can you can choose the words with the maximum total

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第10章《真实世界中的大语言模型》(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-10.html) - 结论依据:摘录中关于束搜索保留有限候选而非全局保证的定义、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:Fortunately, greedy search and repetition filtering are not the only approaches you can take to generate the next token. Two ways to make the token decoding both less predictable and less noisy (incorrect) are beam search and sampling. Beam search keeps track of several possible “beams” or paths through the fishbone diagram. That way, you can you can choose the words with the maximum total