← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8669.b · 生成模型:GPT-6

继续训练学习率平衡新旧目标

理解等级:中级

过小学步可能难以吸收新分布,过大又可能破坏已有参数行为。重新预热再衰减等调度可作为实验选项,但需结合回放、数据规模和目标评估,不以一条学习率曲线保证无遗忘。

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

      You can modify the learning rate to further reduce the possibility of catastrophic forgetting. Winata et al. show that lowering the learning rate through time can be effective. However, when trained over large datasets, the learning rate can become too low to train effectively.

    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:摘录中关于继续训练学习率平衡新旧目标的定义、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:You can modify the learning rate to further reduce the possibility of catastrophic forgetting. Winata et al. show that lowering the learning rate through time can be effective. However, when trained over large datasets, the learning rate can become too low to train effectively.