一、为什么需要微调预训练大模型如GPT、Llama、Qwen具备强大的通用能力但在垂直领域往往表现不佳。微调Fine-tuning通过在特定领域数据上继续训练让模型学会你的业务知识、术语和风格。回到顶部二、LoRA为什么是最佳选择LoRALow-Rank Adaptation的核心思想是冻结预训练模型的原始权重只训练两个低秩矩阵A和B。这将可训练参数量大幅降低成本大幅减少。回到顶部三、环境准备安装依赖pip install transformers peft datasets accelerate bitsandbytes回到顶部四、完整实战代码4.1 数据准备from datasets import load_dataset train_data [ {instruction: 请用Java实现一个线程池, input: , output: ...}, ]4.2 配置LoRAfrom peft import LoraConfig, get_peft_model, TaskType from transformers import AutoModelForCausalLM, AutoTokenizer model AutoModelForCausalLM.from_pretrained(meta-llama/Llama-3.2-3B-Instruct, device_mapauto, load_in_4bitTrue) lora_config LoraConfig( r8, lora_alpha16, target_modules[q_proj, v_proj], lora_dropout0.05, biasnone, task_typeTaskType.CAUSAL_LM ) model get_peft_model(model, lora_config) model.print_trainable_parameters()4.3 训练配置training_args TrainingArguments( output_dir./lora_output, num_train_epochs3, per_device_train_batch_size4, learning_rate2e-4, fp16True, ) trainer Trainer(modelmodel, argstraining_args, train_datasettrain_dataset) trainer.train() model.save_pretrained(./lora_weights)4.4 推理from peft import PeftModel model PeftModel.from_pretrained(base_model, ./lora_weights) model.eval() def chat(prompt): inputs tokenizer(prompt, return_tensorspt).to(cuda) outputs model.generate(**inputs, max_new_tokens512) return tokenizer.decode(outputs[0], skip_special_tokensTrue)回到顶部五、实战效果对比场景微调前LoRA微调后专业术语准确率~45%~92%显存占用28GB~6GB训练时间3B模型数周不可行~4小时