BPE Tokenizer fitted on a custom corpus, with digit separation, byte fallback and other features from LlamaTokenizer.

Only fitted on 1,000,000 samples.