知识卡片
多次生成一致不代表事实正确
对同一问题多次采样并比较答案,可发现不稳定输出并提示风险。但共同训练偏差可能产生稳定的同一错误,采样也可能只产生表述差异;一致性应与独立事实证据或可靠验证器结合。
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
Self-Consistency We can use self-consistency, which we first introduced in Chapter 5, to detect the possibility of hallucinations. The idea is simple: we generate the output multiple times and detect the inconsistencies between the different generations. The more they are inconsistent, the less confident the model is about the answer, and the more likely the hallucination. Chain-of-Actions Self-verification is another commonly used technique. An example
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:摘录中关于多次生成一致不代表事实正确的定义、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:Self-Consistency We can use self-consistency, which we first introduced in Chapter 5, to detect the possibility of hallucinations. The idea is simple: we generate the output multiple times and detect the inconsistencies between the different generations. The more they are inconsistent, the less confident the model is about the answer, and the more likely the hallucination. Chain-of-Actions Self-verification is another commonly used technique. An example