知识卡片

普通读书笔记卡 · 8679 · 生成模型:GPT-6

上下文忠实性不同于事实正确

理解等级:中级

回答可能符合给定段落却违背真实世界,也可能补写段落中没有的信息。上下文任务应分别标记矛盾、无依据添加及外部事实错误;有上下文通常属于open-book设置,原书将其称closed-book不沿用。

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

      In-Context Hallucinations So far, we have focused on hallucinations emanating from the model trying to generate from its parametric memory. We can also have in-context hallucinations, also called closed-book hallucinations, where the model depends on data provided in the context to influence its generation, like in RAG. For example, consider the passage: At the 2020 Summer Olympics in Tokyo, Uganda won three medals: Joshua Cheptegei

    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:摘录中关于上下文忠实性不同于事实正确的定义、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:In-Context Hallucinations So far, we have focused on hallucinations emanating from the model trying to generate from its parametric memory. We can also have in-context hallucinations, also called closed-book hallucinations, where the model depends on data provided in the context to influence its generation, like in RAG. For example, consider the passage: At the 2020 Summer Olympics in Tokyo, Uganda won three medals: Joshua Cheptegei