examples/scripts/sft_gpt_oss.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",
#     "kernels",
#     "trackio",
#     "kernels",
# ]
# ///

"""
pip install –-upgrade kernels

Example:

accelerate launch \
    --config_file examples/accelerate_configs/deepspeed_zero3.yaml \
    examples/scripts/sft_gpt_oss.py \
    --dtype bfloat16 \
    --model_name_or_path openai/gpt-oss-20b \
    --packing \
    --run_name 20b-full-eager \
    --attn_implementation kernels-community/vllm-flash-attn3 \
    --dataset_num_proc 12 \
    --dataset_name HuggingFaceH4/Multilingual-Thinking \
    --max_length 4096 \
    --per_device_train_batch_size 2 \
    --num_train_epochs 1 \
    --logging_steps 1 \
    --warmup_steps 0.03 \
    --lr_scheduler_type cosine_with_min_lr \
    --lr_scheduler_kwargs '{"min_lr_rate": 0.1}' \
    --output_dir gpt-oss-20b-multilingual-reasoner \
    --report_to trackio \
    --seed 42
"""

from datasets import load_dataset
from transformers import AutoModelForCausalLM, Mxfp4Config

from trl import ModelConfig, ScriptArguments, SFTConfig, SFTTrainer, TrlParser, get_peft_config


def main(script_args, training_args, model_args):
    # Load model
    quantization_config = Mxfp4Config(dequantize=True)
    model_kwargs = dict(
        revision=model_args.model_revision,
        trust_remote_code=model_args.trust_remote_code,
        attn_implementation=model_args.attn_implementation,
        dtype=model_args.dtype,
        use_cache=False if training_args.gradient_checkpointing else True,
        quantization_config=quantization_config,
    )

    model = AutoModelForCausalLM.from_pretrained(model_args.model_name_or_path, **model_kwargs)

    # Load dataset
    dataset = load_dataset(script_args.dataset_name, name=script_args.dataset_config)

    # Train model
    trainer = SFTTrainer(
        model=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,
        peft_config=get_peft_config(model_args),
    )

    trainer.train()
    trainer.save_model(training_args.output_dir)
    if training_args.push_to_hub:
        trainer.push_to_hub(dataset_name=script_args.dataset_name)


if __name__ == "__main__":
    parser = TrlParser((ScriptArguments, SFTConfig, ModelConfig))
    script_args, training_args, model_args, _ = parser.parse_args_and_config(return_remaining_strings=True)
    main(script_args, training_args, model_args)