0
0

Qwen2-1.5B-Instruct Lora微调

Qwen2-1.5B-Instruct微调Lora微调

最近做了一个基于Qwen2-1.5B-Instruct模型的比赛，记录一下自己的微调过程。怕自己以后忘了我就手把手一步一步来记录了。
大多数都是给小白看的，如果你是小白建议你用jupyter运行，按照我这个模块一块一块运行，如果你是高手单纯的想找一个训练代码直接看模块10，我在提供了完整代码。

1. 模型下载

一般模型尽量在modelscope上先搜一下，比较这个下载速度真的快。

  1. import torch
  2. from modelscope import snapshot_download, AutoModel, AutoTokenizer
  3. import os
  4. # 第一个参数表示下载模型的型号，第二个参数是下载后存放的缓存地址，第三个表示版本号，默认 master
  5. model_dir = snapshot_download('Qwen/Qwen2-1.5B-Instruct', cache_dir='./', revision='master')

2. 准备工作（高手请忽略，没啥用）

微调的主要工作其实就是数据处理，其他基本就是个架往里放就行。
接下来是一份官网给出的推理的代码，借助这个代码我们来看输入模型的数据格式长什么样。

  1. from modelscope import AutoModelForCausalLM, AutoTokenizer
  2. device ="cuda"# the device to load the model onto
  4. model = AutoModelForCausalLM.from_pretrained("./Qwen2-1.5B-Instruct",
  5.     torch_dtype="auto",
  6.     device_map="auto")
  7. tokenizer = AutoTokenizer.from_pretrained("qwen/Qwen2-1.5B-Instruct")
  9. prompt ="你好"
  10. messages =[{"role":"system","content":'你是医疗问答助手章鱼哥，你将帮助用户解答基础的医疗问题。'},{"role":"user","content": prompt}]
  11. text = tokenizer.apply_chat_template(
  12.     messages,
  13.     tokenize=False,
  14.     add_generation_prompt=True)
  15. model_inputs = tokenizer([text], return_tensors="pt").to(device)
  17. generated_ids = model.generate(
  18.     model_inputs.input_ids,
  19.     max_new_tokens=512)
  20. response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]print(response)

可以打印看看编码后的输入数据长什么样：

  1. '<|im_start|>system\n你是医疗问答助手章鱼哥，你将帮助用户解答基础的医疗问题。<|im_end|>\n<|im_start|>user\n你好<|im_end|>\n<|im_start|>assistant\n'

这里可以看到其实apply_chat_template方法在对函数编码的时候没有给出mask等内容（他这个和智谱轻言的GLM的apply_chat_template就差距很大,在这卡了我半天）所以在数据处理的时候就不能直接用他这个模板。

3. 接下来进入正题吧（导包）

  1. from datasets import Dataset, load_dataset
  2. from transformers import AutoTokenizer, AutoModelForCausalLM, DataCollatorForSeq2Seq, TrainingArguments, Trainer

4. 加载数据

我这里是用了一个医疗问答的csv数据，能了解到这里的应该数据处理不需要细说了吧

  1. dataset = load_dataset("csv", data_files="./问答.csv", split="train")
  2. dataset = dataset.filter(lambda x: x["answer"]isnotNone)
  3. datasets = dataset.train_test_split(test_size=0.1)

5. 数据预处理

  1. tokenizer = AutoTokenizer.from_pretrained("./Qwen2-1.5B-Instruct", trust_remote_code=True)defprocess_func(example):
  2.     MAX_LENGTH =768# 最大输入长度，根据你的显存和数据自己调整
  3.     input_ids, attention_mask, labels =[],[],[]
  4.     instruction = example["question"].strip()# query# instruction = tokenizer.apply_chat_template([{"role": "user", "content": instruction}],#                                    add_generation_prompt=True,#                                    tokenize=True,#                                    )      # '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\nquery<|im_end|>\n<|im_start|>assistant\n'
  5.     instruction = tokenizer(f"<|im_start|>system\n你是医学领域的人工助手章鱼哥<|im_end|>\n<|im_start|>user\n{example['question']}<|im_end|>\n<|im_start|>assistant\n",
  6.         add_special_tokens=False,)
  7.     response = tokenizer(f"{example['answer']}", add_special_tokens=False)# \n response, 缺少eos token
  8.     input_ids = instruction["input_ids"]+ response["input_ids"]+[tokenizer.pad_token_id]
  9.     attention_mask =(instruction["attention_mask"]+ response["attention_mask"]+[1])
  10.     labels =[-100]*len(instruction["input_ids"])+ response["input_ids"]+[tokenizer.pad_token_id]iflen(input_ids)> MAX_LENGTH:
  11.         input_ids = input_ids[:MAX_LENGTH]
  12.         attention_mask = attention_mask[:MAX_LENGTH]
  13.         labels = labels[:MAX_LENGTH]return{"input_ids": input_ids,"attention_mask": attention_mask,"labels": labels
  14.     }
  16. tokenized_ds = datasets['train'].map(process_func, remove_columns=['id','question','answer'])
  17. tokenized_ts = datasets['test'].map(process_func, remove_columns=['id','question','answer'])

6. 创建模型

  1. import torch
  2. model = AutoModelForCausalLM.from_pretrained("./Qwen2-1.5B-Instruct", trust_remote_code=True)from peft import LoraConfig, TaskType, get_peft_model, PeftModel
  4. config = LoraConfig(target_modules=["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"], modules_to_save=["post_attention_layernorm"])# 配置Lora参数
  5. model = get_peft_model(model, config)# 创建Lora模型

7. 配置训练参数

  1. args = TrainingArguments(
  2.     output_dir="./chatbot",
  3.     per_device_train_batch_size=4,
  4.     gradient_accumulation_steps=8,
  5.     gradient_checkpointing=True,
  6.     logging_steps=300,
  7.     num_train_epochs=10,
  8.     learning_rate=1e-4,
  9.     remove_unused_columns=False,
  10.     save_strategy="epoch")# 在这里如果你开起了梯度检查点gradient_checkpointing=True，就必须加上model.enable_input_require_grads()，否则会报一个很难受的错误
  11. model.enable_input_require_grads()

8. 创建训练器

  1. trainer = Trainer(
  2.     model=model,
  3.     args=args,
  4.     train_dataset=tokenized_ds.select(range(5000)),# 我这个数据量很大，我随机抽取5000条训练不然太慢了
  5.     data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True),)

9. 开始训练！！！

祝你成功

  1. trainer.train()

10. 完整的.py代码

  1. import torch
  2. from datasets import Dataset, load_dataset
  3. from transformers import AutoTokenizer, AutoModelForCausalLM, DataCollatorForSeq2Seq, TrainingArguments, Trainer
  4. from peft import LoraConfig, TaskType, get_peft_model, PeftModel
  6. dataset = load_dataset("csv", data_files="./问答.csv", split="train")
  7. dataset = dataset.filter(lambda x: x["answer"]isnotNone)
  8. datasets = dataset.train_test_split(test_size=0.1)
  10. tokenizer = AutoTokenizer.from_pretrained("./Qwen2-1.5B-Instruct", trust_remote_code=True)defprocess_func(example):
  11.     MAX_LENGTH =768
  12.     input_ids, attention_mask, labels =[],[],[]
  13.     instruction = example["question"].strip()# query
  14.     instruction = tokenizer(f"<|im_start|>system\n你是医学领域的人工助手章鱼哥<|im_end|>\n<|im_start|>user\n{example['question']}<|im_end|>\n<|im_start|>assistant\n",
  15.         add_special_tokens=False,)
  16.     response = tokenizer(f"{example['answer']}", add_special_tokens=False)# \n response, 缺少eos token
  17.     input_ids = instruction["input_ids"]+ response["input_ids"]+[tokenizer.pad_token_id]
  18.     attention_mask =(instruction["attention_mask"]+ response["attention_mask"]+[1])
  19.     labels =[-100]*len(instruction["input_ids"])+ response["input_ids"]+[tokenizer.pad_token_id]iflen(input_ids)> MAX_LENGTH:
  20.         input_ids = input_ids[:MAX_LENGTH]
  21.         attention_mask = attention_mask[:MAX_LENGTH]
  22.         labels = labels[:MAX_LENGTH]return{"input_ids": input_ids,"attention_mask": attention_mask,"labels": labels
  23.     }
  25. tokenized_ds = datasets['train'].map(process_func, remove_columns=['id','question','answer'])
  26. tokenized_ts = datasets['test'].map(process_func, remove_columns=['id','question','answer'])
  28. model = AutoModelForCausalLM.from_pretrained("./Qwen2-1.5B-Instruct", trust_remote_code=True)
  30. config = LoraConfig(target_modules=["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"], modules_to_save=["post_attention_layernorm"])
  32. model = get_peft_model(model, config)
  34. args = TrainingArguments(
  35.     output_dir="./law",
  36.     per_device_train_batch_size=4,
  37.     gradient_accumulation_steps=16,
  38.     gradient_checkpointing=True,
  39.     logging_steps=6,
  40.     num_train_epochs=10,
  41.     learning_rate=1e-4,
  42.     remove_unused_columns=False,
  43.     save_strategy="epoch")
  44. model.enable_input_require_grads()
  46. trainer = Trainer(
  47.     model=model,
  48.     args=args,
  49.     train_dataset=tokenized_ds.select(range(400)),
  50.     data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True),)
  52. trainer.train()

11. 合并Lora推理预测代码

  1. import torch
  2. from transformers import AutoModelForCausalLM, AutoTokenizer
  3. from peft import PeftModel
  5. defpredict(messages, model, tokenizer):
  6.     device ="cuda"
  8.     text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
  9.     model_inputs = tokenizer([text], return_tensors="pt").to(device)
  11.     generated_ids = model.generate(model_inputs.input_ids, max_new_tokens=512)
  12.     generated_ids =[output_ids[len(input_ids):]for input_ids, output_ids inzip(model_inputs.input_ids, generated_ids)]
  13.     response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]return response
  15. # 加载原下载路径的tokenizer和model
  16. tokenizer = AutoTokenizer.from_pretrained("./Qwen2-1.5B-Instruct/", use_fast=False, trust_remote_code=True)
  17. model = AutoModelForCausalLM.from_pretrained("./Qwen2-1.5B-Instruct/", device_map="auto", torch_dtype=torch.bfloat16)# 加载训练好的Lora模型，将下面的checkpointXXX替换为实际的checkpoint文件名名称
  18. model = PeftModel.from_pretrained(model, model_id="./chatbot/checkpoint-1560")
  20. test_texts ={'instruction':"你是医学领域的人工助手章鱼哥",'input':"嗓子疼，是不是得了流感了"}
  22. instruction = test_texts['instruction']
  23. input_value = test_texts['input']
  25. messages =[{"role":"system","content":f"{instruction}"},{"role":"user","content":f"{input_value}"}]
  27. response = predict(messages, model, tokenizer)print(response)
标签： 深度学习 人工智能 自然语言处理
本文转载自: https://blog.csdn.net/m0_50972200/article/details/140667653
版权归原作者 十分钟ll 所有， 如有侵权，请联系我们删除。
登录后发布评论

“Qwen2-1.5B-Instruct Lora微调”的评论:

还没有评论
关于作者
...
overfit同步小助手
文章同步
相关阅读

网络安全法-网络运行安全

使用selenium/drissionpage时如何阻止chrome自动跳转http到https

docker desktop 里部署的Open WebUI 管理员密码忘记了的处理方法

在ubuntu20.04中搭建onsite比赛运行环境

利用开源的低代码表单设计器FcDesigner高效管理和渲染复杂表单结构

Kafka学习笔记

【前端】浏览器输入url到页面呈现发生了什么？