← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8678 · 生成模型:GPT-6

预训练分块应保留实体语境

理解等级:中级

将文档切块可能丢失实体、主题或指代背景。给片段添加真实主题标识可帮助保持事实关联,但标识本身需准确,且不能补回所有被切断的跨段关系;应与分块策略及上下文长度共同检查。

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

      Due to the pre-training setup, training documents are often broken into chunks. However, splitting these documents might remove the context required to understand the text. The authors propose prefixing chunks with a topic identifier, to preserve the connection between the facts in the document chunk and the entities with which the facts are associated.

    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:摘录中关于预训练分块应保留实体语境的定义、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:Due to the pre-training setup, training documents are often broken into chunks. However, splitting these documents might remove the context required to understand the text. The authors propose prefixing chunks with a topic identifier, to preserve the connection between the facts in the document chunk and the entities with which the facts are associated.