知识卡片
嵌入噪声只在训练时扰动
训练时向输入嵌入加入受控噪声,可减少对某些措辞或格式的过拟合;推理时通常关闭该噪声。噪声尺度和数据分布影响效果,更长输出或更高词元多样性不自动证明信息更多或能力全面提升。
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
- · 微调
原书出处 · 1 处
- · 查看原书依据
Noise Embeddings The datasets we use for fine-tuning typically consist of a small number of examples (< 50,000). We would like our model to not overfit to the stylistic characteristics of the dataset, like the formatting, wording, and length of the text. One way to address this is by adding noise to the input embeddings. Jain et al. observe that adding noise embeddings reduces the
参考来源
- 位置:《Designing Large Language Model Applications》第06章《微调》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch06.html)
- 结论依据:摘录中关于嵌入噪声只在训练时扰动的定义、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:Noise Embeddings The datasets we use for fine-tuning typically consist of a small number of examples (< 50,000). We would like our model to not overfit to the stylistic characteristics of the dataset, like the formatting, wording, and length of the text. One way to address this is by adding noise to the input embeddings. Jain et al. observe that adding noise embeddings reduces the