How is a training corpus constructed for LLMs$?$

By randomly masking words in the corpus

By training the LLM on pre$-$existing datasets

By predicting the next word without training data

By masking every single word one at a time to create a training data set