from datasets import load_dataset, load_from_disk
from tokenizers import trainers, Tokenizer, normalizers, ByteLevelBPETokenizer
from transformers import AutoConfig

model_dir = "./"  # ${MODEL_DIR}

# load roberta-large config
config = AutoConfig.from_pretrained("roberta-large")
config.save_pretrained(model_dir)

# load dataset
#dataset = load_dataset("oscar", "unshuffled_deduplicated_fi", split="train")
dataset = load_from_disk("DATASET_PATH_HERE")
dataset = dataset["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=config.vocab_size, min_frequency=2, special_tokens=[
    "<s>",
    "<pad>",
    "</s>",
    "<unk>",
    "<mask>",
])

# Save files to disk
tokenizer.save(f"{model_dir}/tokenizer.json")