小白如何开始打AI比赛？记录第一次参加 Datawhale AI夏令营-优快云博客

`内容介绍：`

数学建模竞赛 模型蒸馏 CoT推理链 DeepSeek-R1 轻量化部署

本文记录2025年高等数学推理轻量化竞赛完整解题流程，涵盖数据清洗策略、教师模型知识蒸馏、LoRA微调实战与多维度模型评估，提供可复现的轻量化数学推理解决方案。

优秀的训练集是蒸馏小模型的最好方法，那么我们拿到的数据集，数据质量参差不齐，就需要对数据集进行清洗，减少噪声，具体代码如下

清洗代码

import json
import re

INPUT_FILE = "数学赛题训练集.jsonl"
OUTPUT_FILE = "数学赛题训练集_cleaned.jsonl"

def normalize_text(text: str) -> str:
    # 修正全角标点、空格与引号
    text = text.replace('（', '(').replace('）', ')')
    text = text.replace('，', ',').replace('：', ':')
    text = text.replace('。', '.').replace('！', '!').replace('？', '?')
    text = text.replace('；', ';').replace('“', '"').replace('”', '"')
    text = text.replace('\n', " ").replace('’', "'")
    # 合并多个空格
    text = re.sub(r'\s+', ' ', text).strip()
    return text

def is_valid_entry(entry: dict) -> bool:
    # 必须包含 id, input, output 三个字段且非空
    if not all(k in entry and entry[k].strip() for k in ("id", "input", "output")):
        return False
    return True

def clean_entry(entry: dict) -> dict:
    entry["id"] = entry["id"].strip()
    entry["input"] = normalize_text(entry["input"])
    entry["output"] = normalize_text(entry["output"])
    return entry

def main():
    valid_count = 0
    total = 0
    with open(INPUT_FILE, 'r', encoding='utf-8') as fin, \
         open(OUTPUT_FILE, 'w', encoding='utf-8') as fout:
        for line in fin:
            total += 1
            line = line.strip()
            if not line:
                continue
            try:
                obj = json.loads(line)
            except json.JSONDecodeError:
                # 跳过无法解析的行
                continue
            if not is_valid_entry(obj):
                # 跳过不符合规范的条目
                continue
            obj = clean_entry(obj)
            fout.write(json.dumps(obj, ensure_ascii=False) + "\n")
            valid_count += 1

    print(f"共处理条目: {total}，保留有效条目: {valid_count}")

if __name__ == "__main__":
    main()

~~按照id 进行排序代码，可忽略~~

import json
import os

def sort_jsonl_by_id(input_file, output_file):
    """
    读取JSONL文件，按照id排序，并写入新的JSONL文件
    
    参数:
    input_file -- 输入的JSONL文件路径
    output_file -- 输出的JSONL文件路径
    """
    print(f"正在处理文件: {input_file}")
    
    # 读取JSONL文件
    data = []
    with open(input_file, 'r', encoding='utf-8') as f:
        for line in f:
            try:
                item = json.loads(line.strip())
                data.append(item)
            except json.JSONDecodeError as e:
                print(f"解析JSON时出错: {e}")
                print(f"问题行: {line}")
                continue
    
    print(f"读取了 {len(data)} 条记录")
    
    # 按id排序
    try:
        # 尝试按照数字id排序
        data.sort(key=lambda x: int(x['id']))
        print("按数字id排序成功")
    except (KeyError, ValueError, TypeError):
        try:
            # 如果失败，尝试按照字符串id排序
            data.sort(key=lambda x: str(x.get('id', '')))
            print("按字符串id排序成功")
        except (KeyError, TypeError) as e:
            print(f"排序时出错: {e}")
            print("数据样例:", data[0] if data else "无数据")
            return False
    
    # 写入排序后的数据
    with open(output_file, 'w', encoding='utf-8') as f:
        for item in data:
            f.write(json.dumps(item, ensure_ascii=False) + '\n')
    
    print(f"排序完成，已写入文件: {output_file}")
    return True

def process_all_jsonl_files():
    """处理当前目录下所有的JSONL文件"""
    # 获取当前目录下所有的JSONL文件
    jsonl_files = [f for f in os.listdir('.') if f.endswith('.jsonl')]
    
    print(f"找到 {len(jsonl_files)} 个JSONL文件")
    
    for file in jsonl_files:
        # 生成输出文件名
        base_name = os.path.splitext(file)[0]
        output_file = f"{base_name}_sorted.jsonl"
        
        # 处理文件
        success = sort_jsonl_by_id(file, output_file)
        if success:
            print(f"成功处理文件: {file} -> {output_file}")
        else:
            print(f"处理文件失败: {file}")

def process_specific_files(file_list):
    """处理指定的JSONL文件列表"""
    for file in file_list:
        if not os.path.exists(file):
            print(f"文件不存在: {file}")
            continue
            
        # 生成输出文件名
        base_name = os.path.splitext(file)[0]
        output_file = f"{base_name}_sorted.jsonl"
        
        # 处理文件
        success = sort_jsonl_by_id(file, output_file)
        if success:
            print(f"成功处理文件: {file} -> {output_file}")
        else:
            print(f"处理文件失败: {file}")

if __name__ == "__main__":
    # 指定要处理的文件
    files_to_process = [
        "数学赛题训练集_cleaned.jsonl"
    ]
    
    print("开始按ID排序JSONL文件...")
    process_specific_files(files_to_process)
    print("处理完成!")

数据集优化

噪声解决之后，我们要对文章进行

利用API接口优化数据集，由于平台的训练太慢了，所以智能通过自己的API进行代码优化。

import os
import json
import requests
from tqdm import tqdm

# 配置DeepSeek API
API_KEY = os.getenv('DEEPSEEK_API_KEY', 'sk-fb9dce1a8b5e3c7ad0312af') #key已作废
API_URL = "https://api.deepseek.com/v1/"
MODEL_NAME = "deepseek-r1"

# 上下文长度限制(按token计算)
MAX_CONTEXT_LENGTH = 3000

# 分割长文本的函数
def split_long_text(text, max_length):
    """
    将长文本分割为不超过max_length的片段
    """
    words = text.split()
    segments = []
    current_segment = []
    current_length = 0
    
    for word in words:
        if current_length + len(word) + 1 > max_length:
            segments.append(" ".join(current_segment))
            current_segment = []
            current_length = 0
        
        current_segment.append(word)
        current_length += len(word) + 1
    
    if current_segment:
        segments.append(" ".join(current_segment))
    
    return segments

# 调用DeepSeek API生成CoT推理
def generate_cot_response(prompt):
    headers = {
        "Authorization": f"Bearer {API_KEY}",
        "Content-Type": "application/json"
    }
    
    payload = {
        "model": MODEL_NAME,
        "messages": [
            {"role": "system", "content": "你是一个数学问题解决专家，请给出详细的推理过程(CoT格式)。"},
            {"role": "user", "content": prompt}
        ],
        "temperature": 0.3
    }
    
    try:
        response = requests.post(API_URL, json=payload, headers=headers, timeout=120)
        response.raise_for_status()
        return response.json()["choices"][0]["message"]["content"]
    except Exception as e:
        print(f"API调用失败: {str(e)}")
        return None

# 主处理函数
def process_dataset(input_path, output_path):
    """
    处理JSONL数据集并生成CoT格式输出
    """
    with open(input_path, 'r', encoding='utf-8') as infile, \
         open(output_path, 'w', encoding='utf-8') as outfile:
        
        for line in tqdm(infile, desc="处理进度"):
            try:
                data = json.loads(line)
                question_id = data["id"]
                question_text = data["input"]
                
                # 处理长上下文
                if len(question_text) > MAX_CONTEXT_LENGTH:
                    segments = split_long_text(question_text, MAX_CONTEXT_LENGTH)
                    full_response = ""
                    
                    for segment in segments:
                        response = generate_cot_response(segment)
                        if response:
                            full_response += response + "\n\n"
                else:
                    full_response = generate_cot_response(question_text)
                
                if full_response:
                    result = {
                        "id": question_id,
                        "input": question_text,
                        "output": full_response.strip()
                    }
                    outfile.write(json.dumps(result, ensure_ascii=False) + '\n')
            
            except Exception as e:
                print(f"处理ID {data.get('id', 'unknown')} 时出错: {str(e)}")

if __name__ == "__main__":
    INPUT_PATH = r"d:\Datawhale/数学赛题训练集_cleaned_sorted.jsonl"
    OUTPUT_PATH = r"d:\Datawhale/数学赛题训练集_cot_output.jsonl"
    
    # 检查API密钥
    if API_KEY == "":
        print("请先在代码中设置您的DeepSeek API密钥")
    else:
        process_dataset(INPUT_PATH, OUTPUT_PATH)
        print(f"处理完成! 结果已保存至: {OUTPUT_PATH}")

那么我们得到的数据就一定是好的吗？

不一定，我们还需要对所有的结果进行验证，这样我们导入Kimi 2 只对结果进行验证，来确保训练集中所有的结果都是对的，相当于找了一个外教。