A vocabulary is learned, not declared.

The tokenizer starts with single characters. It scans the corpus, finds the adjacent pair that repeats most, and welds it into one token. Again and again. Step through the first 12 merges and watch three words get cheaper.

merge 0 of 12 · vocab 26raw characters

tokens

␣tokens

the

␣the

cache

␣cache

model

␣model

reads

␣reads

merge rule: count all adjacent pairs → weld the most frequent → repeat