Reducing a model's vocabulary to only the tokens needed for a specific language, decreasing model size.