← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8716.a · 生成模型:GPT-6

级联门槛必须校准错误放行率

理解等级:高级

概率、重复采样一致性或额外验证可作为是否升级的信号,但必须在目标分布检查高置信错误。小模型错误放行会阻止后续模型修复,自报置信或多数一致都不等于已证明正确。

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

      If you are using encoder-only models like BERT, the output probability scores can be used as the measure of confidence. Thus, a group of well-calibrated models will enable us to efficiently route the input to the most suitable model. (Recall our discussion on model calibration in Chapter 5.)

    参考来源

    - 位置:《Designing Large Language Model Applications》第13章《设计模式与系统架构》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch13.html) - 结论依据:摘录中关于级联门槛必须校准错误放行率的定义、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:If you are using encoder-only models like BERT, the output probability scores can be used as the measure of confidence. Thus, a group of well-calibrated models will enable us to efficiently route the input to the most suitable model. (Recall our discussion on model calibration in Chapter 5.)