知识卡片
学习率须与目标和配置联合调节
学习率影响每次参数更新尺度,其合适范围依模型、可训练参数、数据、优化器和批量。原书推荐数值是示例起点,不能不经验证当通用默认;调参还应保留稳定任务样本和需保留能力。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 微调
原书出处 · 1 处
- · 查看原书依据
Learning rates For each optimizer, certain learning rates have been shown to be very effective. A recommended learning rate for AdamW is 1e-4 with a weight decay of 0.01. Weight decay is a regularization technique that helps reduce overfitting. Similarly, the default values for minor optimizer parameters like adam_beta1, adam_beta2, and adam_epsilon are good enough and need not be changed. Exercise Learning rate rules for
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:摘录中关于学习率须与目标和配置联合调节的定义、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:Learning rates For each optimizer, certain learning rates have been shown to be very effective. A recommended learning rate for AdamW is 1e-4 with a weight decay of 0.01. Weight decay is a regularization technique that helps reduce overfitting. Similarly, the default values for minor optimizer parameters like adam_beta1, adam_beta2, and adam_epsilon are good enough and need not be changed. Exercise Learning rate rules for