trl/scripts/reward.py

# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# /// script
# dependencies = [
#     "trl",
#     "peft",
#     "trackio",
#     "kernels",
# ]
# ///

import argparse


def main(script_args, training_args, model_args, dataset_args):
    from accelerate import logging
    from datasets import load_dataset

    from trl import RewardTrainer, get_dataset, get_peft_config

    logger = logging.get_logger(__name__)

    # Load the dataset
    if dataset_args.datasets and script_args.dataset_name:
        logger.warning(
            "Both `datasets` and `dataset_name` are provided. The `datasets` argument will be used to load the "
            "dataset and `dataset_name` will be ignored."
        )
        dataset = get_dataset(dataset_args)
    elif dataset_args.datasets and not script_args.dataset_name:
        dataset = get_dataset(dataset_args)
    elif not dataset_args.datasets and script_args.dataset_name:
        dataset = load_dataset(
            script_args.dataset_name, name=script_args.dataset_config, streaming=script_args.dataset_streaming
        )
    else:
        raise ValueError("Either `datasets` or `dataset_name` must be provided.")

    # Initialize the RewardTrainer
    trainer = RewardTrainer(
        model=model_args.model_name_or_path,
        args=training_args,
        train_dataset=dataset[script_args.dataset_train_split],
        eval_dataset=dataset[script_args.dataset_test_split] if training_args.eval_strategy != "no" else None,
        peft_config=get_peft_config(model_args),
    )

    # Train the model
    trainer.train()

    # Log training complete
    trainer.accelerator.print("✅ Training completed.")

    # Save and push to Hub
    trainer.save_model(training_args.output_dir)
    trainer.accelerator.print(f"💾 Model saved to {training_args.output_dir}.")

    if training_args.push_to_hub:
        trainer.push_to_hub(dataset_name=script_args.dataset_name)
        trainer.accelerator.print(f"🤗 Model pushed to the Hub in https://huggingface.co/{trainer.hub_model_id}.")


def make_parser(subparsers: argparse._SubParsersAction | None = None, prog: str | None = None):
    from trl import DatasetMixtureConfig, ModelConfig, RewardConfig, ScriptArguments, TrlParser

    dataclass_types = (ScriptArguments, RewardConfig, ModelConfig, DatasetMixtureConfig)
    if subparsers is not None:
        parser = subparsers.add_parser(
            "reward", help="Run the reward training script", dataclass_types=dataclass_types
        )
    else:
        parser = TrlParser(dataclass_types, prog=prog)
    return parser


if __name__ == "__main__":
    parser = make_parser()
    script_args, training_args, model_args, dataset_args = parser.parse_args_and_config(fail_with_unknown_args=False)
    main(script_args, training_args, model_args, dataset_args)