← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8670.e · 生成模型:GPT-6

微调能否获得新能力要可检验

理解等级:高级

某些微调主要改善已有能力的调用方式,另一些可学习新的映射或行为。要论证“新能力”需先定义能力、基准和训练信息;不能由部分研究概括为微调永远只暴露旧能力或不能吸收新知识。

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

      Does Fine-Tuning Learn New Capabilities? This is an important question with heavy implications. There is increasing evidence that fine-tuning (the way it is performed today) only exposes already existing capabilities and doesn’t necessarily impart new capabilities. If this is the case, then one can find a subset of parameters that is more amenable to solving a given downstream task. Zhao et al. propose using a

    参考来源

    - 位置:《Designing Large Language Model Applications》第07章《进阶微调技术》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch07.html) - 结论依据:摘录中关于微调能否获得新能力要可检验的定义、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:Does Fine-Tuning Learn New Capabilities? This is an important question with heavy implications. There is increasing evidence that fine-tuning (the way it is performed today) only exposes already existing capabilities and doesn’t necessarily impart new capabilities. If this is the case, then one can find a subset of parameters that is more amenable to solving a given downstream task. Zhao et al. propose using a