一位淘宝工程同学的大模型LoRA微调尝试

pip install --upgrade torch==2.1.0+cpu torchvision==0.16.0+cpu --index-url https://download.pytorch.org/whl/cpupip install -U transformers==4.38.2 peft==0.10.0 datasets==2.18.0 \            accelerate sentencepiece safetensors tqdm

{"instrucAon":"埋点时主要记录哪些事件？","response":"使⽤的是xx埋点框架，埋点有分为各种事件类型，主要的有曝光事件和点击事件，可以根据曝光和点击事件来统计数据，如点击率等"}{"instrucAon":"xxxxxxxxxxxxxxxxxx？","response":"xxxxxxxxxxxxxxx"}

import jsonfrom datasets import Datasetfrom transformers import AutoTokenizer# 正确读取jsonlwith open("lora_seckill_qa.jsonl", "r", encoding="utf-8") as f:raw_data = [json.loads(line) for line in f if line.strip()]# 如果是单轮格式可直接用# dataset = Dataset.from_list(raw_data)# 若是conversation 格式（如 [{"conversation":[...]}]），需展开def conversation_to_list(item):out = []for turn in item["conversation"]:instr = turn.get("system", "") + "\n" + turn["input"] ifturn.get("system") else turn["input"]out.append({"instruction": instr.strip(),"response": turn["output"].strip()})return out# 如果你的raw_data 已经是单轮格式就跳过这一段all_samples = []if "conversation" in raw_data[0]:for d in raw_data:all_samples.extend(conversation_to_list(d))dataset = Dataset.from_list(all_samples)else:dataset = Dataset.from_list(raw_data)
# 保存分词器tokenizer = AutoTokenizer.from_pretrained("/Users/shawn/Documents/AI-dev/models/deepseek/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",trust_remote_code=True,local_files_only=True)tokenizer.save_pretrained("./tokenizer")# 保存处理后的数据集dataset.save_to_disk("./processed_dataset_ms")print("预处理完成，已保存！")

import torchfrom datasets import load_from_diskfrom transformers import AutoTokenizer, AutoModelForCausalLM,TrainingArguments, Trainer, default_data_collatorfrom peft import get_peft_model, LoraConfigimport random# 1. 加载数据集dataset = load_from_disk("./processed_dataset_ms")# 2. 样本随机打乱dataset = dataset.shuffle(seed=42)# 3. 分词器tokenizer = AutoTokenizer.from_pretrained("./tokenizer", local_files_only=True)def generate_and_tokenize_prompt(batch):texts = [f"""<s>### Instruction:{instruction}### Response:{response}</s>"""for instruction, response in zip(batch["instruction"],batch["response"])]out = tokenizer(texts,max_length=256,padding="max_length",truncation=True,add_special_tokens=False,return_tensors=None,)# Loss 忽略paddingp.out["labels"] = [[tok if tok != tokenizer.pad_token_id else -100 for tok in label]for label in out["input_ids"]]return outtokenized_dataset = dataset.map(generate_and_tokenize_prompt,batched=True,remove_columns=dataset.column_names,desc="Tokenizing")# 4. 加载基座模型（如设备有限可指定CPU/其他device）model = AutoModelForCausalLM.from_pretrained("/Users/shawn/Documents/AI-dev/models/deepseek/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",torch_dtype=torch.bfloat16, # MPS、A100 等建议用bfloat16local_files_only=True,trust_remote_code=True)# 5. LoRA 配置lora_config = LoraConfig(r=8,lora_alpha=16,target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],lora_dropout=0.05, # 或0, 稳定死记记忆；推荐小数据降dropoutbias="none",task_type="CAUSAL_LM")model = get_peft_model(model, lora_config)model.print_trainable_parameters()# 6. 训练参数training_args = TrainingArguments(output_dir="./results_ms",per_device_train_batch_size=2,gradient_accumulation_steps=1,learning_rate=3e-4,num_train_epochs=8,logging_dir="./logs_ms",save_steps=100,save_total_limit=3,logging_steps=10,overwrite_output_dir=True,report_to=None,fp16=False)trainer = Trainer(model=model,args=training_args,train_dataset=tokenized_dataset,data_collator=default_data_collator,)# 7. 训练与自动采样输出trainer.train()# 8. 保存模型与tokenizermodel.save_pretrained("./deepseek-7b-lora_ms")tokenizer.save_pretrained("./deepseek-7b-lora_ms")print("训练完成。")

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel# 路径设置BASE_PATH = "/Users/shawn/Documents/AI-dev/models/deepseek/deepseekai/DeepSeek-R1-Distill-Qwen-7B"LORA_PATH = "./deepseek-7b-lora_ms"# 加载分词器tokenizer = AutoTokenizer.from_pretrained(BASE_PATH,local_files_only=True,trust_remote_code=True)# === 严格分离模型实例 ===# 原始模型model_base = AutoModelForCausalLM.from_pretrained(BASE_PATH,device_map="mps",torch_dtype=torch.bfloat16,local_files_only=True,trust_remote_code=True)# 微调后的模型（Base + LoRA 适配器）model_base_for_lora = AutoModelForCausalLM.from_pretrained(BASE_PATH,device_map="mps",torch_dtype=torch.bfloat16,local_files_only=True,trust_remote_code=True)model_lora = PeftModel.from_pretrained(model_base_for_lora,LORA_PATH)def format_prompt_one_round(user_input: str) -> str:return f"<s>### Instruction:\n{user_input}\n### Response:\n"def generate_single(model, prompt, tokenizer, max_new_tokens=200):inputs = tokenizer(prompt, return_tensors="pt")for k, v in inputs.items():inputs[k] = v.to(model.device)with torch.no_grad():outputs = model.generate(**inputs,max_new_tokens=max_new_tokens,temperature=0.7,top_p=0.9,do_sample=True,eos_token_id=tokenizer.eos_token_id)full_output = tokenizer.decode(outputs[0], skip_special_tokens=True)reply = full_output[len(prompt):]# 截断下一个分隔符，保证只输出新生成内容for sep in ["<s>", "</s>", "###"]:if sep in reply:reply = reply.split(sep)[0]return reply.strip()def main():print("="*40)print("DeepSeek 模型微调前/后 单轮对话对比（严格模型物理分离版）")print("="*40)print("输入exit 退出。\n")while True:user_input = input("你说：").strip()if user_input.lower() in {"exit", "quit"}:print("对话结束~ 再见！")breakprompt = format_prompt_one_round(user_input)# 原始模型推理base_reply = generate_single(model_base, prompt, tokenizer)# LoRA 微调后模型推理lora_reply = generate_single(model_lora, prompt, tokenizer)print("\n-------------------------------")print("【原模型 输出】↓\n" + base_reply)print("\n【微调后输出】↓\n" + lora_reply)print("-------------------------------")if __name__ == "__main__":main()

{“messages": [{"role": "system", "content": "You are a professional e-commercetitle analyst. Given a long product title, output ONLY the core product entityname (1-5 words) without any other text."}, {"role": "user", "content": "请从下列商品标题中提取最核心的商品主体，直接输出主体名，不要加其它词：【黑旗】心语肉松原味辣味1kg 烘焙面包蛋糕寿司原料商用肉松小贝"}, {"role": "assistant", "content": "肉松"}]}{"messages": [{"role": "system", "content": "You are a professional e-commercetitle analyst. Given a long product title, output ONLY the core product entityname (1-5 words) without any other text."}, {"role": "user", "content": "请从下列商品标题中提取最核心的商品主体，直接输出主体名，不要加其它词：夏季竹枕片成人藤凉席冰丝枕头套单人儿童竹枕席藤枕芯套一对拍2"}, {"role": "assistant", "content": "枕套"}]}