xtuner/examples/huggingface_trainer/train_hf.py

# Copyright (c) OpenMMLab. All rights reserved.
import transformers
from transformers import Trainer

from xtuner.apis import DefaultTrainingArguments, build_model
from xtuner.apis.datasets import alpaca_data_collator, alpaca_dataset


def train():
    # get DefaultTrainingArguments and to be updated with passed args
    parser = transformers.HfArgumentParser(DefaultTrainingArguments)
    training_args = parser.parse_args_into_dataclasses()[0]

    # init model and dataset
    model, tokenizer = build_model(
        model_name_or_path=training_args.model_name_or_path,
        return_tokenizer=True)
    train_dataset = alpaca_dataset(
        tokenizer=tokenizer, path=training_args.dataset_name_or_path)
    data_collator = alpaca_data_collator(return_hf_format=True)

    # build trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=train_dataset,
        data_collator=data_collator)

    # training
    trainer.train()

    trainer.save_state()
    trainer.save_model(output_dir=training_args.output_dir)


if __name__ == '__main__':
    train()