from datasets import load_dataset from tokenizers import trainers, Tokenizer, normalizers, ByteLevelBPETokenizer model_dir = "." # ${MODEL_DIR} # load dataset dataset = load_dataset("allegro_reviews", split="train") # Instantiate tokenizer tokenizer = ByteLevelBPETokenizer() def batch_iterator(batch_size=1000): for i in range(0, len(dataset), batch_size): yield dataset[i: i + batch_size]["text"] # Customized training tokenizer.train_from_iterator(batch_iterator(), vocab_size=50265, min_frequency=2, special_tokens=[ "", "", "", "", "", ]) # Save files to disk tokenizer.save(f"{model_dir}/tokenizer.json")