知识卡片
适配器初始行为应接近基座
新增模块若从一开始就大幅改变信息流,可能扰乱已有模型。近恒等或小增量初始化使起点更接近基础行为,再让训练学习任务变化;它帮助平滑开始,不是后续不回退的保证。
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
- · 进阶微调技术
原书出处 · 1 处
- · 查看原书依据
Therefore, a bottleneck adapter sublayer consists of a down-projection matrix, an up-projection matrix at the end to project back to the original dimensions, and parameters that can be configured in a variety of ways in the middle. During fine-tuning, only the adapter modules are updated. The original pre-trained model is not updated. Adapters are initialized with a near-identity initialization to ensure smooth training.
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:摘录中关于适配器初始行为应接近基座的定义、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:Therefore, a bottleneck adapter sublayer consists of a down-projection matrix, an up-projection matrix at the end to project back to the original dimensions, and parameters that can be configured in a variety of ways in the middle. During fine-tuning, only the adapter modules are updated. The original pre-trained model is not updated. Adapters are initialized with a near-identity initialization to ensure smooth training.