知识卡片

普通读书笔记卡 · 8712.b · 生成模型:GPT-6

动态示例以任务收益衡量

理解等级:高级

相似输入可帮助选择few-shot例子,但主题相似不一定最有教学价值。可在训练数据上比较加入示例后金标输出似然,学习排序或检索器;在线使用时只能依赖已学规则与可用反馈,不应泄漏测试答案。

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

      RAG for Selecting In-Context Training Examples As mentioned at the beginning of the chapter, another application of RAG is to dynamically select training examples for few-shot learning by retrieving the optimal examples from a data store containing a list of training examples. For a given input, the retrieved few-shot examples are supposed to maximize the LLM’s chance of generating the correct answer to a user

    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:摘录中关于动态示例以任务收益衡量的定义、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:RAG for Selecting In-Context Training Examples As mentioned at the beginning of the chapter, another application of RAG is to dynamically select training examples for few-shot learning by retrieving the optimal examples from a data store containing a list of training examples. For a given input, the retrieved few-shot examples are supposed to maximize the LLM’s chance of generating the correct answer to a user