Trim a tokenizer's vocabulary, and optionally its model's embedding table, down to the subset you actually need. You can specify "what you need" explicitly, by presets, or by providing a corpus (one or more datasets) as an anchor for what you'd actually like to keep.
