知识卡片
删除高风险来源牺牲部分覆盖
当某文档或来源敏感内容密度高时,整段或整源排除可避免逐项处理的复杂性,也会损失其有效知识。阈值应依据风险与任务覆盖,记录删除原因及影响,不能只以处理方便替代内容判断。
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
Remove entire document/data source If the amount of PII detected in a single document or data source is higher than a specific threshold, it is probably best to remove it. For example, pastebin.com is said to contain a lot of inadvertently placed PII and is recommended to be not included in training datasets. Each of these techniques can have a varied effect on the model’s
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:摘录中关于删除高风险来源牺牲部分覆盖的定义、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:Remove entire document/data source If the amount of PII detected in a single document or data source is higher than a specific threshold, it is probably best to remove it. For example, pastebin.com is said to contain a lot of inadvertently placed PII and is recommended to be not included in training datasets. Each of these techniques can have a varied effect on the model’s