知识卡片
合成数据需检查覆盖与错误传播
模型可生成样例或辅助标注,减少部分人工工作,但会带入生成偏差与错误。需保留来源、抽查标签并在真实独立样例上比较收益;数量更多不意味着覆盖或公平必然改善。
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
原书出处 · 1 处
- · 查看原书依据
In the age of LLMs, it has become very common to use them to generate additional natural language data. Using LLMs for everything from generating samples of a particular user intent to creating questions and answers for a Q&A dataset is how many teams nowadays mitigate a lack of real-world data. And it’s not uncommon to use LLMs to generate labels too, by evaluating the
参考来源
- 位置:《Natural Language Processing in Action, Second Edition》附录D《附录D:机器学习工具与技术》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/appendix-d.html)
- 结论依据:摘录中关于合成数据需检查覆盖与错误传播的定义、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:In the age of LLMs, it has become very common to use them to generate additional natural language data. Using LLMs for everything from generating samples of a particular user intent to creating questions and answers for a Q&A dataset is how many teams nowadays mitigate a lack of real-world data. And it’s not uncommon to use LLMs to generate labels too, by evaluating the