开源模型结构与微调实战：以 Qwen3-4B-Instruct-2507 为例

Qwen3-4B-Instruct-2507/
├── README.md                        # 模型说明文档（model card）
│                                     # 包含模型简介、使用方法、性能指标
├── LICENSE                          # 许可证（Apache 2.0）
│                                     # 允许商用与二次开发，需保留版权声明
├── config.json                      # 模型结构配置
│                                     # 定义 Transformer 层数、隐藏维度、注意力头数等参数
├── generation_config.json           # 推理参数配置
│                                     # 控制文本生成行为（如 temperature、top_p、max_new_tokens）
├── tokenizer_config.json            # 分词器参数定义
│                                     # 控制是否小写化、padding、特殊 token 映射等
├── tokenizer.json                   # 分词器核心文件
│                                     # 包含词汇表与 BPE 规则，定义 token <-> id 映射
├── vocab.json                       # 词汇表
│                                     # 每个 token 对应的整数 ID
├── merges.txt                       # BPE 合并规则
│                                     # 描述字符合并为 token 的规则
├── model.safetensors.index.json     # 模型权重索引文件
│                                     # 指示权重被拆分到哪些 .safetensors 文件
├── model-00001-of-00003.safetensors # 模型权重文件（分片 1）
├── model-00002-of-00003.safetensors # 模型权重文件（分片 2）
├── model-00003-of-00003.safetensors # 模型权重文件（分片 3）
│                                     # safetensors 格式加载更安全高效
└── .gitattributes                   # Git 属性定义，用于 LFS/xet 管理大文件

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "Qwen/Qwen3-4B-Instruct-2507"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

prompt = "请简要解释一下大模型微调（Fine-tuning）的基本原理。"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=300,
    temperature=0.7,
    top_p=0.9
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

conda create -n qwen3 python=3.10 -y
conda activate qwen3

pip install torch --index-url https://download.pytorch.org/whl/cu121
pip install transformers accelerate bitsandbytes peft datasets trl

#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
Qwen3-4B-Instruct-2507 LoRA/QLoRA 微调脚本
"""
import os
import argparse
import torch
from datasets import load_dataset
from transformers import (
    AutoTokenizer, AutoModelForCausalLM,
    Trainer, TrainingArguments,
    DataCollatorForLanguageModeling,
    BitsAndBytesConfig,
)
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training

def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument("--dataset_path", type=str, required=True)
    parser.add_argument("--output_dir", type=str, default="./output-lora")
    parser.add_argument("--use_qlora", action="store_true")
    return parser.parse_args()

def load_dataset_tokenized(path, tokenizer):
    data = load_dataset("json", data_files=path)
    def preprocess(example):
        prompt = f"指令：{example['instruction']}\n输入：{example.get('input','')}\n回答：{example['output']}"
        return tokenizer(prompt, truncation=True, max_length=1024, padding="max_length")
    return data["train"].map(preprocess)

def main():
    args = parse_args()
    model_name = "Qwen/Qwen3-4B-Instruct-2507"

    tokenizer = AutoTokenizer.from_pretrained(model_name)
    tokenizer.pad_token = tokenizer.eos_token

    if args.use_qlora:
        bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16)
        model = AutoModelForCausalLM.from_pretrained(model_name, quantization_config=bnb, device_map="auto", trust_remote_code=True)
        model = prepare_model_for_kbit_training(model)
    else:
        model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, device_map="auto", trust_remote_code=True)

    lora_cfg = LoraConfig(r=8, lora_alpha=32, lora_dropout=0.05,
                          target_modules=["q_proj", "v_proj"], task_type="CAUSAL_LM")
    model = get_peft_model(model, lora_cfg)
    dataset = load_dataset_tokenized(args.dataset_path, tokenizer)
    data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)

    args_train = TrainingArguments(
        output_dir=args.output_dir, per_device_train_batch_size=2,
        gradient_accumulation_steps=8, num_train_epochs=3, learning_rate=2e-4,
        fp16=not args.use_qlora, bf16=args.use_qlora, save_strategy="epoch"
    )

    trainer = Trainer(model=model, args=args_train, train_dataset=dataset,
                      tokenizer=tokenizer, data_collator=data_collator)
    trainer.train()
    trainer.save_model(args.output_dir)
    tokenizer.save_pretrained(args.output_dir)

if __name__ == "__main__":
    main()

[
  {"instruction": "解释什么是云原生", "input": "", "output": "云原生是一种基于容器、微服务和 Kubernetes 的现代软件架构思想。"},
  {"instruction": "翻译以下句子", "input": "AI is changing software engineering.", "output": "AI 正在改变软件工程。"}
]

python finetune_qwen.py --dataset_path ./train.json --output_dir ./output-lora

python finetune_qwen.py --dataset_path ./train.json --use_qlora

from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("./output-lora", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("./output-lora")

prompt = "请解释 Kubernetes 和 Docker 的关系。"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

output-lora/
├── adapter_config.json        # LoRA 配置文件
├── adapter_model.bin          # LoRA 权重
├── config.json                # 模型结构配置（继承原模型）
├── tokenizer.json             # 分词器配置
├── tokenizer_config.json
└── training_args.bin          # 训练参数