← 返回卡片

原书出处

Designing Large Language Model Applications

First Edition · 核验版本的摘录快照 · 导入于 2026-11-14T13:54:13.942Z

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

此预览只包含定位所需的摘录和上下文。

原书定位详情

来源:a19c7933-4eda-4176-aebd-30fe7e1ad292

正文位置:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch08.html

EPUB SHA-256:1b4d2347fec2f949fcddfbaf4fad597a1ae20887a096cfae64c59a619b4781bf