How to train a new language model from scratch using Transformers and Tokenizers