Does anyone know if tokenizers are pruned? That is, if a token doesn't appear in the corpus is it removed from the model? That would imply a process that leaks

Does anyone know if tokenizers are pruned? That is, if a token doesn't appear in the corpus is it removed from the model? That would imply a process that leaks information about the dataset.