知识卡片
质量标准会偏向参考分布
按token分布与参考语料的KL偏离来筛选,本质上偏好与该参考相近的文档。偏离可能是噪声,也可能是少数领域或新词;需要说明分布估计、平滑和阈值,不能把统计距离自动解释成事实质量。
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
Token-distribution K-L divergence In this method, documents with a token distribution that deviates too much from a reference token distribution are removed. In effect, this removes documents that have a lot of outlier tokens. This is calculated by using the Kullback-Liebler (K-L) divergence. Classifier-based approaches We can also build a classifier for identifying high-quality data. A simple way to build a quality-based classifier is to
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:摘录中关于质量标准会偏向参考分布的定义、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:Token-distribution K-L divergence In this method, documents with a token distribution that deviates too much from a reference token distribution are removed. In effect, this removes documents that have a lot of outlier tokens. This is calculated by using the Kullback-Liebler (K-L) divergence. Classifier-based approaches We can also build a classifier for identifying high-quality data. A simple way to build a quality-based classifier is to