知识卡片
置信信号必须经过校准
低输出置信可作为风险信号,但token概率或整句似然不自动等于答案正确概率。要在目标任务按置信分桶比较实际准确率,并检查对齐后是否变化;仅存在正相关还不足以证明校准良好。
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
assess a model’s self-knowledge is through its output uncertainty. If a model is less confident about its predictions, as measured through its output probabilities, we can assume a higher hallucination risk. For this approach to be valid, the model has to be well calibrated. As Chapter 6 introduced, a model is well calibrated if there is a correlation between its output probability values and task
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:摘录中关于置信信号必须经过校准的定义、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:assess a model’s self-knowledge is through its output uncertainty. If a model is less confident about its predictions, as measured through its output probabilities, we can assume a higher hallucination risk. For this approach to be valid, the model has to be well calibrated. As Chapter 6 introduced, a model is well calibrated if there is a correlation between its output probability values and task