知识卡片
排序蒸馏继承教师和候选偏差
用初检候选和大模型排序训练较小排序器,可降低运行成本。打乱候选输入顺序可检查位置依赖,仍须验证教师错误与候选覆盖;初检很强时可提升空间可能变小。
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
- · 检索增强生成
原书出处 · 1 处
- · 查看原书依据
Examples of ranking LLMs include RankGPT, RankVicuna, and RankZephyr. These models are trained by distilling from larger LLMs, a technique we first learned in Chapter 9. For example, the process for training RankVicuna is: Queries in the training set are fed through a first-level retriever like BM25 to generate a list of candidate documents. This list is passed to a larger LLM, which generates a
参考来源
- 位置:《Designing Large Language Model Applications》第12章《检索增强生成》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch12.html)
- 结论依据:摘录中关于排序蒸馏继承教师和候选偏差的定义、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:Examples of ranking LLMs include RankGPT, RankVicuna, and RankZephyr. These models are trained by distilling from larger LLMs, a technique we first learned in Chapter 9. For example, the process for training RankVicuna is: Queries in the training set are fed through a first-level retriever like BM25 to generate a list of candidate documents. This list is passed to a larger LLM, which generates a