知识卡片

普通读书笔记卡 · 8667 · 生成模型:GPT-6

先生成输入可能造成标签失衡

理解等级:中级

从指令先生成输入再标输出,容易集中于教师更常见的类别。也可先指定标签再构造满足它的输入以控制覆盖;两种方式都需人工或规则核对,不能只因指定了标签就认定样本符合。

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

      input-first approach, where the LLM is asked to generate an input instance for the given instruction and subsequently asked to generate the output label for that input. However, this approach might lead to label imbalance as shown in Wang et al., with certain labels being overrepresented. Therefore, it is a good approach to mix output-first generation, where you ask the LLM to generate the output

    参考来源

    - 位置:《Designing Large Language Model Applications》第06章《微调》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch06.html) - 结论依据:摘录中关于先生成输入可能造成标签失衡的定义、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:input-first approach, where the LLM is asked to generate an input instance for the given instruction and subsequently asked to generate the output label for that input. However, this approach might lead to label imbalance as shown in Wang et al., with certain labels being overrepresented. Therefore, it is a good approach to mix output-first generation, where you ask the LLM to generate the output