trl/scripts/kto.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",
# ]
# ///

"""
Run the KTO training script with the commands below. In general, the optimal configuration for KTO will be similar to
that of DPO.

# Full training:
```bash
python trl/scripts/kto.py \
    --dataset_name trl-lib/kto-mix-14k \
    --model_name_or_path=trl-lib/qwen1.5-1.8b-sft \
    --per_device_train_batch_size 16 \
    --num_train_epochs 1 \
    --learning_rate 5e-7 \
    --lr_scheduler_type=cosine \
    --gradient_accumulation_steps 1 \
    --eval_steps 500 \
    --output_dir=kto-aligned-model \
    --warmup_steps 0.1 \
    --logging_first_step
```

# QLoRA:
```bash
# QLoRA:
python trl/scripts/kto.py \
    --dataset_name trl-lib/kto-mix-14k \
    --model_name_or_path=trl-lib/qwen1.5-1.8b-sft \
    --per_device_train_batch_size 8 \
    --num_train_epochs 1 \
    --learning_rate 5e-7 \
    --lr_scheduler_type=cosine \
    --gradient_accumulation_steps 1 \
    --eval_steps 500 \
    --output_dir=kto-aligned-model-lora \
    --warmup_steps 0.1 \
    --logging_first_step \
    --use_peft \
    --load_in_4bit \
    --lora_target_modules=all-linear \
    --lora_r=16 \
    --lora_alpha=16
```
"""

import argparse


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

    from trl import get_dataset, get_peft_config
    from trl.experimental.kto import KTOTrainer

    logger = logging.get_logger(__name__)

    # Load a pretrained model
    model = AutoModelForCausalLM.from_pretrained(
        model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code
    )
    ref_model = AutoModelForCausalLM.from_pretrained(
        model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code
    )

    tokenizer = AutoTokenizer.from_pretrained(
        model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code
    )
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token

    # 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 KTO trainer
    trainer = KTOTrainer(
        model,
        ref_model,
        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,
        processing_class=tokenizer,
        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, ScriptArguments, TrlParser
    from trl.experimental.kto import KTOConfig

    dataclass_types = (ScriptArguments, KTOConfig, ModelConfig, DatasetMixtureConfig)
    if subparsers is not None:
        parser = subparsers.add_parser("kto", help="Run the KTO 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)