知识卡片

普通读书笔记卡 · 8700 · 生成模型:GPT-6

领域适配应定义想改变的距离

理解等级:中级

微调嵌入可让任务相关正例更近、负例更远,例如区分主题相同但极性不同的句子。改变某种关系可能损害其他任务,需要保持集和独立验证;不能把空间调整想成所有维度都同时变好。

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

      similarity dimensions are important to us. If negation is an important relation for our application to distinguish, it might be a good idea to reflect that in our embedding space. This is where fine-tuning embedding models can come in handy. Fine-tuning embedding models allows you to “edit” your embedding space to your own liking. The process is relatively simple and can be potentially quite beneficial.

    参考来源

    - 位置:《Designing Large Language Model Applications》第11章《表示学习与嵌入》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch11.html) - 结论依据:摘录中关于领域适配应定义想改变的距离的定义、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:similarity dimensions are important to us. If negation is an important relation for our application to distinguish, it might be a good idea to reflect that in our embedding space. This is where fine-tuning embedding models can come in handy. Fine-tuning embedding models allows you to “edit” your embedding space to your own liking. The process is relatively simple and can be potentially quite beneficial.