trl + 大模型reward训练

一、定义

1.强化学习-reward训练
2.reward 模型重新加载与训练

二、实现

https://www.kaggle.com/code/neuqsnail/open-llama-finetune-sequenceclassification/notebook#Save-and-reload-Model
1.trl 强化训练-reward训练案例
#注意：lora训练需要 task_type 为 SEQ_CLS
1. 下载trl 训练脚本
2. 指令训练

python examples/scripts/reward_modeling.py \
    --model_name_or_path Qwen/Qwen2-0.5B-Instruct \
    --dataset_name trl-lib/ultrafeedback_binarized \
    --output_dir Qwen2-0.5B-Reward-LoRA \
    --per_device_train_batch_size 8 \
    --num_train_epochs 1 \
    --gradient_checkpointing True \
    --learning_rate 1.0e-4 \
    --logging_steps 25 \
    --eval_strategy steps \
    --eval_steps 50 \
    --max_length 2048 \
    --task_type SEQ_CLS\
    --use_peft \
    --lora_r 32 \
    --lora_alpha 16

对应代码

import warnings

import torch
from datasets import load_dataset
from transformers import AutoModelForSequenceClassification, AutoTokenizer, HfArgumentParser

from trl import (
    ModelConfig,
    RewardConfig,
    RewardTrainer,
    ScriptArguments,
    get_kbit_device_map,
    get_peft_config,
    get_quantization_config,
    setup_chat_format,
)


if __name__ == "__main__":
    parser = HfArgumentParser((ScriptArguments, RewardConfig, ModelConfig))
    script_args, training_args, model_args = parser.parse_args_into_dataclasses()
    training_args.gradient_checkpointing_kwargs = dict(use_reentrant=False)
        
    ################
    # Model & Tokenizer
    ################
    torch_dtype = (
        model_args.torch_dtype if model_args.torch_dtype in ["auto", None] else getattr(torch, model_args.torch_dtype)
    )
    quantization_config = get_quantization_config(model_args)
    model_kwargs = dict(
        revision=model_args.model_revision,
        device_map=get_kbit_device_map() if quantization_config is not None else None,
        quantization_config=quantization_config,
        use_cache=False if training_args.gradient_checkpointing else True,
        torch_dtype=torch_dtype,
    )
    tokenizer = AutoTokenizer.from_pretrained(
        model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code, use_fast=True
    )
    model = AutoModelForSequenceClassification.from_pretrained(
        model_args.model_name_or_path, num_labels=1, trust_remote_code=model_args.trust_remote_code, **model_kwargs
    )
    # Align padding tokens between tokenizer and model
    model.config.pad_token_id = tokenizer.pad_token_id

    # If post-training a base model, use ChatML as the default template
    if tokenizer.chat_template is None:
        model, tokenizer = setup_chat_format(model, tokenizer)

    if model_args.use_peft and model_args.lora_task_type != "SEQ_CLS":
        warnings.warn(
            "You are using a `task_type` that is different than `SEQ_CLS` for PEFT. This will lead to silent bugs"
            " Make sure to pass --lora_task_type SEQ_CLS when using this script with PEFT.",
            UserWarning,
            #reward 应该是SEQ_CLS 类型
        )

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

    ##########
    # Training
    ##########
    trainer = RewardTrainer(
        model=model,
        processing_class=tokenizer,
        args=training_args,
        train_dataset=dataset[script_args.dataset_train_split].select(range(50)),
        eval_dataset=dataset[script_args.dataset_test_split].select(range(50)) if training_args.eval_strategy != "no" else None,
        peft_config=get_peft_config(model_args),
    )
    trainer.train()

    ############################
    # Save model and push to Hub
    ############################
    trainer.save_model(training_args.output_dir)

     

    if training_args.eval_strategy != "no":
        metrics = trainer.evaluate()
        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)

    # Save and push to hub
    trainer.save_model(training_args.output_dir)
    if training_args.push_to_hub:
        trainer.push_to_hub(dataset_name=script_args.dataset_name)

#输入：数据处理: 需要字段chosen、rejected 两个字段
目标是： 对两者进行估分，并使选中的概率大于拒绝的概率，越大越好。

def _tokenize(batch: dict[str, list[Any]], tokenizer: "PreTrainedTokenizerBase") -> dict[str, list[Any]]:
    """Tokenize a batch from a reward modelling dataset."""
    new_examples = {
        "input_ids_chosen": [],
        "attention_mask_chosen": [],
        "input_ids_rejected": [],
        "attention_mask_rejected": [],
    }
    for chosen, rejected in zip(batch["chosen"], batch["rejected"]):
        tokenized_chosen = tokenizer(chosen)
        tokenized_rejected = tokenizer(rejected)
        new_examples["input_ids_chosen"].append(tokenized_chosen["input_ids"])
        new_examples["attention_mask_chosen"].append(tokenized_chosen["attention_mask"])
        new_examples["input_ids_rejected"].append(tokenized_rejected["input_ids"])
        new_examples["attention_mask_rejected"].append(tokenized_rejected["attention_mask"])
    return new_examples

#损失函数

rewards_chosen = model(
            input_ids=inputs["input_ids_chosen"],
            attention_mask=inputs["attention_mask_chosen"],
            return_dict=True,
        )["logits"]
        rewards_rejected = model(
            input_ids=inputs["input_ids_rejected"],
            attention_mask=inputs["attention_mask_rejected"],
            return_dict=True,
        )["logits"]
loss = -nn.functional.logsigmoid(rewards_chosen - rewards_rejected).mean()

2.模型重新加载

from peft import AutoPeftModelForSequenceClassification
from transformers import AutoTokenizer
import torch

adapter_model_name = "Qwen2-0.5B-Reward-LoRA"
model = AutoPeftModelForSequenceClassification.from_pretrained("Qwen2-0.5B-Reward-LoRA", num_labels=1)
#和训练保持一致
tokenizer = AutoTokenizer.from_pretrained(
        "E:\\Qwen2.5-0.5B-Instruct",  use_fast=True
    )
# Align padding tokens between tokenizer and model  注意需要和训练保持一致
model.config.pad_token_id = tokenizer.pad_token_id
print(model)