DPO with WinRateCallback

dpo_winrate.py

# flake8: noqa
# Copyright 2023 The HuggingFace Inc. 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.
"""
# Full training
python examples/scripts/dpo.py \
    --dataset_name trl-lib/ultrafeedback_binarized \
    --model_name_or_path Qwen/Qwen2-0.5B-Instruct \
    --learning_rate 5.0e-7 \
    --num_train_epochs 1 \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 8 \
    --gradient_checkpointing \
    --logging_steps 25 \
    --eval_strategy steps \
    --eval_steps 1 \
    --output_dir Qwen2-0.5B-DPO \
    --no_remove_unused_columns --dataset_num_proc 12

# LoRA:
python examples/scripts/dpo.py \
    --dataset_name trl-lib/ultrafeedback_binarized \
    --model_name_or_path Qwen/Qwen2-0.5B-Instruct \
    --learning_rate 5.0e-6 \
    --num_train_epochs 1 \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 8 \
    --gradient_checkpointing \
    --logging_steps 25 \
    --eval_strategy steps \
    --eval_steps 50 \
    --output_dir Qwen2-0.5B-DPO \
    --no_remove_unused_columns \
    --use_peft \
    --lora_r 32 \
    --lora_alpha 16
"""

from trl.commands.cli_utils import DPOScriptArguments, TrlParser
from trl.trainer.utils import SIMPLE_CHAT_TEMPLATE
import torch
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer
from accelerate import PartialState
from trl import (
    DPOConfig,
    DPOTrainer,
    ModelConfig,
    get_kbit_device_map,
    get_peft_config,
    get_quantization_config,
    maybe_extract_prompt,
    maybe_apply_chat_template,
)


if __name__ == "__main__":
    parser = TrlParser((DPOScriptArguments, DPOConfig, ModelConfig))
    args, training_args, model_config = parser.parse_args_and_config()

    ################
    # Model & Tokenizer
    ###################
    torch_dtype = (
        model_config.torch_dtype
        if model_config.torch_dtype in ["auto", None]
        else getattr(torch, model_config.torch_dtype)
    )
    quantization_config = get_quantization_config(model_config)
    model_kwargs = dict(
        revision=model_config.model_revision,
        attn_implementation=model_config.attn_implementation,
        torch_dtype=torch_dtype,
        use_cache=False if training_args.gradient_checkpointing else True,
        device_map=get_kbit_device_map() if quantization_config is not None else None,
        quantization_config=quantization_config,
    )
    model = AutoModelForCausalLM.from_pretrained(
        model_config.model_name_or_path, trust_remote_code=model_config.trust_remote_code, **model_kwargs
    )
    peft_config = get_peft_config(model_config)
    if peft_config is None:
        ref_model = AutoModelForCausalLM.from_pretrained(
            model_config.model_name_or_path, trust_remote_code=model_config.trust_remote_code, **model_kwargs
        )
    else:
        ref_model = None
    tokenizer = AutoTokenizer.from_pretrained(
        model_config.model_name_or_path, trust_remote_code=model_config.trust_remote_code
    )
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    if tokenizer.chat_template is None:
        tokenizer.chat_template = SIMPLE_CHAT_TEMPLATE
    if args.ignore_bias_buffers:
        # torch distributed hack
        model._ddp_params_and_buffers_to_ignore = [
            name for name, buffer in model.named_buffers() if buffer.dtype == torch.bool
        ]

    ################
    # Dataset
    ################
    dataset = load_dataset(args.dataset_name)

    with PartialState().local_main_process_first():
        dataset = dataset.map(maybe_extract_prompt, num_proc=training_args.dataset_num_proc)
        dataset = dataset.map(
            maybe_apply_chat_template, num_proc=training_args.dataset_num_proc, fn_kwargs={"tokenizer": tokenizer}
        )

    ##########
    # Training
    ################
    trainer = DPOTrainer(
        model,
        ref_model,
        args=training_args,
        train_dataset=dataset[args.dataset_train_split],
        eval_dataset=dataset[args.dataset_test_split],
        tokenizer=tokenizer,
        peft_config=peft_config,
    )
    from trl.trainer.judges import OpenAIPairwiseJudge
    from trl import WinRateCallback
    from transformers import GenerationConfig

    gen_config = GenerationConfig(max_new_tokens=1024, do_sample=True, temperature=1.0)
    judge = OpenAIPairwiseJudge(model="gpt-4o-mini")
    cbk = WinRateCallback(judge=judge, trainer=trainer, generation_config=gen_config, num_prompts=100)
    trainer.add_callback(cbk)

    trainer.train()
    metrics = trainer.evaluate()
    trainer.log_metrics("eval", metrics)
    trainer.save_metrics("eval", metrics)
    trainer.save_model(training_args.output_dir)