知识卡片
护栏统计不能替代真实漏检评估
过滤分数和命中日志可帮助排查与抽样,但只观察已拦截内容看不到漏检。需用独立标注与持续采样估计误报、漏报及漂移;增加训练例也不保证各类指标单调改善,过滤器仍有自己的误差。
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
Your filters have another invaluable feature that an LLM cannot provide: statistical measures of how well your LLM pipeline is doing. Your analytics platform will be able to keep track of all the times your LLM came close to saying something that would have exceeded your bad behavior thresholds. In a production system, it is impossible to read all the things your chatbot and users
参考来源
- 位置:《Natural Language Processing in Action, Second Edition》第10章《真实世界中的大语言模型》(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-10.html)
- 结论依据:摘录中关于护栏统计不能替代真实漏检评估的定义、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:Your filters have another invaluable feature that an LLM cannot provide: statistical measures of how well your LLM pipeline is doing. Your analytics platform will be able to keep track of all the times your LLM came close to saying something that would have exceeded your bad behavior thresholds. In a production system, it is impossible to read all the things your chatbot and users