← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8714.a · 生成模型:GPT-6

长窗口评估要加入相关干扰

理解等级:中级

单一随机needle配无关背景主要检验定位回忆,不充分覆盖多事实整合和相关干扰。应加入相似实体、单位变化、跨段条件及需组合证据的真实问题,并观察长度增长时质量与成本。

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

      needle-in-a-haystack tests have been devised. One such implementation by Greg Kamradt facilitates adding a random fact or statement (the needle) to the middle of the context (the haystack) and then asking the LLM questions for which the needle is the answer.

    参考来源

    - 位置:《Designing Large Language Model Applications》第12章《检索增强生成》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch12.html) - 结论依据:摘录中关于长窗口评估要加入相关干扰的定义、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:needle-in-a-haystack tests have been devised. One such implementation by Greg Kamradt facilitates adding a random fact or statement (the needle) to the middle of the context (the haystack) and then asking the LLM questions for which the needle is the answer.