知识卡片
回放混合保留旧分布信号
将旧数据与新数据混合,可在更新过程中继续提供原能力训练信号。比例和覆盖影响新旧取舍,旧数据不可获得时也有限制;原书25%到80%的新数据比例是作者经验,不是通用配方。
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
Replay (Memory) Replay-based techniques are one of the simplest techniques to alleviate catastrophic forgetting. In this approach, we store pre-training examples from the original dataset and interleave them with the continual training dataset. Thus, the data drift is not so pronounced. The following formula has worked very well for me: sample from different subsets of the original pre-training datasets and mix them with the continual
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:摘录中关于回放混合保留旧分布信号的定义、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:Replay (Memory) Replay-based techniques are one of the simplest techniques to alleviate catastrophic forgetting. In this approach, we store pre-training examples from the original dataset and interleave them with the continual training dataset. Thus, the data drift is not so pronounced. The following formula has worked very well for me: sample from different subsets of the original pre-training datasets and mix them with the continual