知识卡片
合成数据由种子与提示塑形
合成预训练数据常从课程大纲、主题或网页片段出发,用提示生成扩写内容。种子覆盖、教师能力、生成约束和过滤决定最终分布;大规模生成不自动带来新事实或消除教师错误。
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
is the use of LLMs to generate synthetic data that can be used for pre-training LLMs. One of the first success stories in training LLMs on datasets with a significant proportion of synthetic data is Microsoft’s phi series of models. For the phi-1.5 model, Microsoft created 20 billion tokens of synthetic data, using 20,000 seed topics and samples from real-world web datasets in their prompts.
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:摘录中关于合成数据由种子与提示塑形的定义、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:is the use of LLMs to generate synthetic data that can be used for pre-training LLMs. One of the first success stories in training LLMs on datasets with a significant proportion of synthetic data is Microsoft’s phi series of models. For the phi-1.5 model, Microsoft created 20 billion tokens of synthetic data, using 20,000 seed topics and samples from real-world web datasets in their prompts.