知识卡片
幻觉评估先声明事实参照
本章将无训练或输入依据的生成称为幻觉,但实际通常无法逐条检查模型训练全集。产品评估应明确采用外部事实准确性、给定上下文忠实性或两者,避免把检索文档中的错误复述为“有依据所以正确”。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 对齐训练与推理
原书出处 · 1 处
- · 查看原书依据
Colloquially, the word “hallucination” is used to refer to instances where the LLM generates something deemed to be incorrect. But LLMs can be incorrect for a wide variety of reasons. In this book, we will lead with a narrower definition of hallucinations: “An LLM is said to have hallucinated a generation if the generated text is not grounded to its training data or input context.”
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:摘录中关于幻觉评估先声明事实参照的定义、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:Colloquially, the word “hallucination” is used to refer to instances where the LLM generates something deemed to be incorrect. But LLMs can be incorrect for a wide variety of reasons. In this book, we will lead with a narrower definition of hallucinations: “An LLM is said to have hallucinated a generation if the generated text is not grounded to its training data or input context.”