使用modelScope微调deepseek

从modelScope中下载模型：

modelscope download --model deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B

在home下创建项目名：deepseek_finetune

创建目录结构：

再增加：根目录下创建 output

在data下再创建： processed_data

run.sh:

#!/bin/bash

# 数据预处理
python3 scripts/preprocess.py

# 启动训练（单卡）
CUDA_VISIBLE_DEVICES=0 python3 scripts/train.py

# 多卡训练示例（需安装 accelerate）
# accelerate launch --num_processes 2 scripts/train.py

config/train_config.json:

{
    "model_id": "deepseek-ai/DeepSeekDeepSeek-R1-Distill-Qwen-1___5B",

  }
  

data/train.jsonl:

{
    "conversation": [
        {"human": "如何学习Python？", "assistant": "可以从基础语法开始，推荐《Python编程：从入门到实践》"},
        {"human": "世界上最伟大的球员是谁？", "assistant": "克里斯蒂亚诺"},
        {"human": "你是谁？", "assistant": "我的名字是：小爱助手"}
    ]
}

scripts/preprocess.py

from datasets import load_dataset
from modelscope import AutoModelForCausalLM, AutoTokenizer
import os

# 使用本地路径加载模型和分词器
model_path = "/home/yygh/.cache/modelscope/hub/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-1___5B"
tokenizer = AutoTokenizer.from_pretrained(model_path)

# 加载数据集
dataset = load_dataset("json", data_files="/home/deepseek_finetune/data/train.jsonl", split="train")

# 确保输出目录存在
output_dir = "/home/deepseek_finetune/data/processed_data"
if not os.path.exists(output_dir):
    os.makedirs(output_dir)

# 定义预处理函数
def preprocess_function(examples):
    texts = []
    # 检查 examples 是否是字典类型，并打印每个元素
    for example in examples:
        # 打印 example 来查看结构
        print(example)
        # 确保 example 是字典类型
        if isinstance(example, dict):
            conversation = example.get("conversation", [])
            if isinstance(conversation, list):
                conversation_text = " ".join([turn.get("human", "") + " " + turn.get("assistant", "") for turn in conversation])
                texts.append(conversation_text)
            else:
                texts.append("")  # 如果 conversation 格式不对，添加空字符串
        else:
            texts.append("")  # 如果 example 不是字典类型，添加空字符串
    return tokenizer(texts, truncation=True, max_length=512)

# 应用预处理
tokenized_data = dataset.map(preprocess_function, batched=True)

# 保存处理后的数据
tokenized_data.save_to_disk(output_dir)

scripts/train.py:

import json
import torch
import numpy as np
from datasets import load_from_disk
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    TrainingArguments,
    Trainer,
    BitsAndBytesConfig,
    DataCollatorForLanguageModeling
)
from peft import (
    LoraConfig,
    get_peft_model,
    prepare_model_for_kbit_training
)



# ---- 2. 量化配置 ----
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16
)

# ---- 3. 模型和分词器加载 ----
model_path = "/home/yygh/.cache/modelscope/hub/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-1___5B"
tokenizer = AutoTokenizer.from_pretrained(model_path)
tokenizer.pad_token = tokenizer.eos_token  # 确保有填充token

model = AutoModelForCausalLM.from_pretrained(
    model_path,
    quantization_config=bnb_config,
    device_map="auto",
    torch_dtype=torch.float16
)

# ---- 4. LoRA配置 ----
lora_config = LoraConfig(
    r=8,
    lora_alpha=32,
    target_modules=["q_proj", "v_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

# ---- 5. 准备模型 ----
model = prepare_model_for_kbit_training(model)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()  # 修正方法名

# ---- 6. 数据加载与处理 ----
dataset = load_from_disk("/home/deepseek_finetune/data/processed_data")



    
    
    

# 确保数据格式正确
def format_dataset(example):
    return {
        "input_ids": np.array(example["input_ids"]),
        "attention_mask": np.array(example["attention_mask"]),
        "labels": np.array(example["input_ids"])  # 复制为labels
    }

dataset = dataset.map(
    format_dataset,
    num_proc=4 if len(dataset) > 1 else 1,
    remove_columns=[col for col in dataset.column_names if col not in ["input_ids", "attention_mask", "labels"]]
)

# ---- 7. 数据整理器 ----
collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer,
    mlm=False  # 因果语言模型
)

# ---- 8. 训练参数 ----
training_args = TrainingArguments(
    output_dir="/home/deepseek_finetune/output",
    per_device_train_batch_size=4,          # 保守起见的batch_size
    gradient_accumulation_steps=8,
    learning_rate=2e-4,
    num_train_epochs=3,
    fp16=True,
    logging_steps=50,
    save_steps=500,
    optim="adamw_torch_fused",
    report_to="tensorboard",
    remove_unused_columns=True,             # 自动过滤无用列
    warmup_ratio=0.1,
    lr_scheduler_type="cosine",
    evaluation_strategy="steps" if "validation" in dataset else "no",
    dataloader_num_workers=4,
    label_names=["labels"]                  # 显式声明标签字段
)

# ---- 9. 训练器 ----
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=dataset,
    data_collator=collator,
    eval_dataset=dataset["validation"] if "validation" in dataset else None
)

# ---- 10. 启动训练 ----
print("===== 开始训练 =====")
print(f"样本数: {len(dataset)}, 批次大小: {training_args.per_device_train_batch_size}")
print(f"可训练参数占比: {model.print_trainable_parameters()}")

trainer.train()

# ---- 11. 保存结果 ----
model.save_pretrained("/home/deepseek_finetune/output/lora")
tokenizer.save_pretrained("/home/deepseek_finetune/output/lora")
print("训练完成，模型已保存")

根目录下执行命令：