知识卡片
偏好反馈有不同信息粒度
接受或拒绝提供单个结果判断,成对比较给出相对偏好,排序表达多个候选次序,纠正反馈则提供理想答案或改法。收集时应区分偏好与事实正确,保留任务上下文及标注标准,不能只把格式不同的数据无条件混用。
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
Binary feedback In this setting, the feedback is provided as either yes/no (accept/reject). Binary comparisons In this setting, the human evaluates outputs A and B and specifies their preference among the two. Ranking In this setting, the human evaluates a set of outputs and provides a rank ordering of preferences. Corrective feedback In this setting, the human explicitly states what should have been the ideal
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:摘录中关于偏好反馈有不同信息粒度的定义、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:Binary feedback In this setting, the feedback is provided as either yes/no (accept/reject). Binary comparisons In this setting, the human evaluates outputs A and B and specifies their preference among the two. Ranking In this setting, the human evaluates a set of outputs and provides a rank ordering of preferences. Corrective feedback In this setting, the human explicitly states what should have been the ideal