知识卡片

普通读书笔记卡 · 8674.e · 生成模型:GPT-6

偏好样本必须共享比较上下文

理解等级:中级

chosen与rejected表达在同一请求及对话背景下的相对选择。要明确哪部分为共有提示、哪部分为待比较回答,避免把不同历史对话误配成偏好对;被选答案也不意味着内容全面正确。

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

      chosen and one rejected. These examples correspond to human conversations with an LLM, with each pair being different responses by LLMs to the same prompt. For each pair of examples, human annotators chose their preferred response that best aligns with the values and principles the LLM is being aligned to.

    参考来源

    - 位置:《Designing Large Language Model Applications》第08章《对齐训练与推理》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch08.html) - 结论依据:摘录中关于偏好样本必须共享比较上下文的定义、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:chosen and one rejected. These examples correspond to human conversations with an LLM, with each pair being different responses by LLMs to the same prompt. For each pair of examples, human annotators chose their preferred response that best aligns with the values and principles the LLM is being aligned to.