知识卡片
层间对比分布突出晚层变化
DoLa将中间层投影到词表后的分布与最终层对比,用相对变化重新构造解码分数。候选层与模型规模影响效果,不应把“低层只学语法、高层只存事实”当严格分工,也不保证模型不知道的事实凭空出现。
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
- · 对齐训练与推理
原书出处 · 1 处
- · 查看原书依据
decoding by contrasting layers (DoLa) is that factual knowledge is encoded in the topmost layer of the Transformer, just like syntactic information is encoded in the lower layers. Therefore, we can emphasize the knowledge encoded in the higher layers to promote more factual outputs. DoLa achieves this by using a technique called contrastive decoding, in which the next token probability for each token is calculated
参考来源
- 位置:《Designing Large Language Model Applications》第08章《对齐训练与推理》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch08.html)
- 结论依据:摘录中关于层间对比分布突出晚层变化的定义、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:decoding by contrasting layers (DoLa) is that factual knowledge is encoded in the topmost layer of the Transformer, just like syntactic information is encoded in the lower layers. Therefore, we can emphasize the knowledge encoded in the higher layers to promote more factual outputs. DoLa achieves this by using a technique called contrastive decoding, in which the next token probability for each token is calculated