fine_tune_improved.py

import torch
from datasets import load_dataset
from peft import LoraConfig, get_peft_model
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from trl import SFTTrainer

# 1. 加载模型和分词器
# 使用Qwen2.5-0.5B模型，小尺寸适合CPU训练
model_name = "Qwen/Qwen2.5-0.5B"  # 使用小尺寸的Qwen模型

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float32,
    low_cpu_mem_usage=True,
    trust_remote_code=True,
)
model.config.use_cache = False

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
# 为Qwen添加pad token
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# 2. 加载并准备数据集
dataset = load_dataset("json", data_files="data.json", split="train")

# 将数据集转换为文本格式，添加结束标记
def convert_to_text(examples):
    texts = []
    for i in range(len(examples['instruction'])):
        text = f"### Instruction:\n{examples['instruction'][i]}\n\n### Input:\n{examples['input'][i]}\n\n### Response:\n{examples['output'][i]}{tokenizer.eos_token}"
        texts.append(text)
    return {"text": texts}

dataset = dataset.map(convert_to_text, batched=True, remove_columns=dataset.column_names)

# 3. 配置LoRA参数
lora_config = LoraConfig(
    r=16,  # 增加rank
    lora_alpha=32,
    lora_dropout=0.1,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],  # Qwen的注意力和MLP模块
)

# 4. 创建PEFT模型
model = get_peft_model(model, lora_config)

# 5. 配置训练参数
output_dir = "./qwen2.5-0.5b-lora"
training_args = TrainingArguments(
    output_dir=output_dir,
    per_device_train_batch_size=2,  # 稍微增加批次大小
    gradient_accumulation_steps=4,   # 减少梯度累积
    learning_rate=2e-4,             # 调整学习率
    logging_steps=10,
    max_steps=200,                  # 增加训练步数
    save_strategy="steps",
    save_steps=50,
    dataloader_num_workers=0,
    fp16=False,
    report_to=[],
    remove_unused_columns=False,
    warmup_steps=20,                # 添加预热步数
    weight_decay=0.01,              # 添加权重衰减
)

# 6. 创建SFTTrainer
trainer = SFTTrainer(
    model=model,
    train_dataset=dataset,
    args=training_args,
    peft_config=lora_config,
    max_seq_length=256,  # 限制序列长度
)

# 7. 开始训练
print("开始训练Qwen2.5-0.5B模型...")
trainer.train()

# 8. 保存模型
print("保存Qwen2.5-0.5B LoRA适配器...")
trainer.save_model(output_dir)
print(f"Qwen2.5-0.5B LoRA适配器已保存到 {output_dir}")