A sliding window moves a fixed-size span across tokenized text, producing local sequences that overlap as the window advances. It turns one long stream into many model-ready examples. section 14.5.4 section 16.2
Nine-token CBOW windows
With a window size of nine and a context size of four, each window contains four tokens on the left, one middle token, and four tokens on the right. Remove the middle token to form an eight-token context bag, and use it as the label to predict the missing word. Advancing the window by one token repeats this process for the rest of the text. section 14.5.4







