Ah ok, I think we made different assumptions about whether the model was specific to the particular dataset so each one would need a new model — a dictionary is specific to the particular dataset being compressed, right? I was thinking the LLM would be a general-purpose text compression model.