How do you mathematically represent and handle out-of-vocabulary tokens in modern tokenizer implementations for Trunk Tools?
