知识卡片

普通读书笔记卡 · 8710.b · 生成模型:GPT-6

紧耦合摘要用答案收益筛选

理解等级:高级

可用带金标训练样本,比较候选摘要加入后正确答案的条件似然,挑选有利摘要训练压缩器。该目标优化下游回答而非人类可读性,可能依赖特定生成器;换模型后需重新评估。

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

      While you can leverage the LLM’s zero-shot capabilities for both extractive and abstractive summarization, it is more effective (albeit expensive) to fine-tune them so that the summaries generated are specifically optimized to enable the LLM to generate the correct answer. We will call these tightly-coupled summarizers.

    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:摘录中关于紧耦合摘要用答案收益筛选的定义、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:While you can leverage the LLM’s zero-shot capabilities for both extractive and abstractive summarization, it is more effective (albeit expensive) to fine-tune them so that the summaries generated are specifically optimized to enable the LLM to generate the correct answer. We will call these tightly-coupled summarizers.