原书出处
Designing Large Language Model Applications
In the MLM setting, rather than predict the next token in a sequence, we ask the model to predict masked tokens within the sequence. In the most basic form of MLM implemented in the BERT model, 15% of tokens are randomly chosen to be masked and are replaced with a special mask token, and the language model is asked to predict the original tokens.
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/ch04.html
EPUB SHA-256:1b4d2347fec2f949fcddfbaf4fad597a1ae20887a096cfae64c59a619b4781bf