知识卡片
重采样要有预算和有效验收器
生成多个候选并筛选可改善部分输出,前提是验收器能识别目标错误。应限制次数、延迟及费用,失败后回退;提高温度可能扩大错误,意图分数通过也不保证数学结果或事实成立。
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
原书出处 · 1 处
- · 查看原书依据
In chapter 10, you saw several ways LLMs can go off the rails and generate unhelpful, incorrect, or even toxic messages. That’s why you never want to use LLMs directly, without any grounding, fine-tuning, or guardrails. It’s better to combine them with other techniques—for example, you can use intent recognition to flag any user messages that might trigger a toxic reply from an LLM. And
参考来源
- 位置:《Natural Language Processing in Action, Second Edition》第12章《对话引擎与会话设计》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-12.html)
- 结论依据:摘录中关于重采样要有预算和有效验收器的定义、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 chapter 10, you saw several ways LLMs can go off the rails and generate unhelpful, incorrect, or even toxic messages. That’s why you never want to use LLMs directly, without any grounding, fine-tuning, or guardrails. It’s better to combine them with other techniques—for example, you can use intent recognition to flag any user messages that might trigger a toxic reply from an LLM. And