知识卡片
去重与提高采样权重不矛盾
去重移除意外冗余,重采样则按训练目标有意改变某些来源的权重。应分别记录唯一内容量、重复副本和实际见到的token次数;否则会把训练预算增长误当新增知识覆盖。
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
- · 预训练数据
原书出处 · 1 处
- · 查看原书依据
Duplicate documents are removed during the preparation of most pre-training datasets. However, in some datasets like The Pile, certain subsets (like Wikipedia) are deliberately duplicated, so that they are seen more often by the model.
参考来源
- 位置:《Designing Large Language Model Applications》第02章《预训练数据》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch02.html)
- 结论依据:摘录中关于去重与提高采样权重不矛盾的定义、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:Duplicate documents are removed during the preparation of most pre-training datasets. However, in some datasets like The Pile, certain subsets (like Wikipedia) are deliberately duplicated, so that they are seen more often by the model.