知识卡片
困惑度的指数底数须匹配
若平均负对数似然用自然对数计算,PPL=exp(loss);若用log₂,则PPL=2^loss。两种底数不可混用。还需固定tokenizer、计分目标和有效长度,才能比较困惑度。
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
- · 架构与学习目标
原书出处 · 1 处
- · 查看原书依据
Perplexity = 2^Cross-Entropy Transformer Backbones So far, we described the components of the canonical version of the Transformer. In practice, three major types of architecture backbones are used to implement the Transformer: Encoder-only Encoder-decoder Decoder-only Let’s look at each of these in detail. Figure 4-5 depicts encoder-only, encoder-decoder, and decoder-only architectures. Figure 4-5. Visualization of various Transformer backbones Encoder-Only Architectures Encoder-only architectures were all the
参考来源
- 位置:《Designing Large Language Model Applications》第04章《架构与学习目标》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch04.html)
- 结论依据:摘录中关于困惑度的指数底数须匹配的定义、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:Perplexity = 2^Cross-Entropy Transformer Backbones So far, we described the components of the canonical version of the Transformer. In practice, three major types of architecture backbones are used to implement the Transformer: Encoder-only Encoder-decoder Decoder-only Let’s look at each of these in detail. Figure 4-5 depicts encoder-only, encoder-decoder, and decoder-only architectures. Figure 4-5. Visualization of various Transformer backbones Encoder-Only Architectures Encoder-only architectures were all the