顶会论文复现：PROVING TEST SET CONTAMINATION IN BLACK BOX LANGUAGE MODELS

1 资料

我复现的源码:https://github.com/Whiffe/test_set_contamination

官网源码：https://github.com/tatsu-lab/test_set_contamination

论文：https://openreview.net/forum?id=KS8mIvetg2

论文翻译：ICLR-2024.Oren.PROVING TEST SET CONTAMINATION IN BLACK BOX LANGUAGE MODELS

b站复现视频：https://www.bilibili.com/video/BV14d1CYWE26/

2 我的总结

这篇论文的测试数据污染的方法也是很扯淡的，论文结尾也说了，作者自己的方法得先证明数据集内的题目之间的顺序打乱是否有影响，这不就是扯淡么，训练期间，有个策略就是要每次输入训练时，打乱顺序，训练的时候都打乱了，作者测试期间打乱测的出来个屁呀。这也能发顶会，太离谱了。

还有检测时用的logprobs这个值，这个值的低和高不代表污染程度，整个论文让我感到匪夷所思。

我依旧对作者的源码进行了重构，简单化地跑了以下，不然按照作者的本地化部署，巨量的遍历循环数据集，那得跑到啥时候，即使能等它跑完，对大模型的消耗也是巨大的，我用的api来跑，那可是按量收费的。

3 复现源码

首先你需要有gpt的api接口

# 设置API key和API的基础URL，用于调用 OpenAI 接口
API_KEY = ""  # 替换为你的 API key
BASE_URL = ""  # 替换为API的基本URL

安装：

pip install fire
pip install openai

数据集

同一个数据集被打乱

执行指令

python main.py --dataset_path "benchmarks/boolq/dev2.jsonl" --log_file_path "results_with_qwen_logprobs.json"

源码

'''
python main.py --dataset_path "benchmarks/boolq/dev2.jsonl" --log_file_path "results_with_qwen_logprobs.json"

此代码用于加载数据集、通过 OpenAI API 计算 token 序列的 logprobs，并通过统计检验比较原始顺序和打乱顺序的 logprobs。
'''

# 导入所需的库和模块
import os
import math
import random 
import numpy as np  # 数组和数值计算库
from scipy.stats import t as tdist  # 导入t分布用于统计检验
from multiprocessing import Process, Queue  # 用于并行处理
from tqdm import tqdm  # 用于显示进度条
import json  # 处理 JSON 数据
import fire  # 用于命令行接口
from openai import OpenAI  # 导入OpenAI库用于调用GPT模型

# 设置API key和API的基础URL，用于调用 OpenAI 接口
API_KEY = ""  # 替换为你的 API key
BASE_URL = ""  # 替换为API的基本URL

# 创建OpenAI客户端，用于后续调用API
client = OpenAI(api_key=API_KEY, base_url=BASE_URL)

# 定义两个 lambda 函数：用于展平嵌套列表和打乱列表
flatten = lambda l : [x for s in l for x in s]  # 展平嵌套列表
shuffle = lambda l : random.sample(l, k=len(l))  # 打乱列表

def load_dataset(dataset_path):
    # 加载数据集函数
    if dataset_path.endswith(".json"):
        # 如果是JSON文件，读取内容
        print("loading from json...")
        with open(dataset_path, "r") as f:
            data = f.read()
            examples = json.loads(data)  # 将JSON格式数据解析为Python对象
            return examples

    # 如果不是JSON，逐行读取文件
    with open(dataset_path, "r") as f:
        lines = f.readlines()  # 读取所有行
    return lines  # 返回行列表

def compute_logprob_of_token_sequence(tokens, context_len=2048, device=0):
    """
    调用 OpenAI API 计算一系列 token 的对数概率 (logprobs)。
    """
    # 将token列表合并成一个输入字符串
    input_text = " ".join(tokens)

    try:
        # 使用 GPT 模型调用API并请求返回logprobs
        response = client.chat.completions.create(
            messages=[{"role": "user", "content": input_text}],
            model="gpt-3.5-turbo",  # 使用GPT-3.5模型
            logprobs=True  # 请求返回 logprobs
        )
        
        # 从响应中提取 logprobs
        logprobs = [token_logprob.logprob for token_logprob in response.choices[0].logprobs.content]
        
        # 计算并返回所有 token 的 logprobs 的和
        total_logprob = sum(logprobs)

        return total_logprob  # 返回 logprobs 总和

    except Exception as e:
        # 如果发生错误，打印错误信息
        print(f"An error occurred: {e}")
        return None  # 返回 None 以表示失败

def worker(context_len, device, main_queue, worker_queue):
    # 工作进程，用于处理多个并行任务
    while True:
        # 从 worker_queue 获取 token 列表、shard ID 和是否是 canonical（原始顺序）
        tokens, shard_id, is_canonical = worker_queue.get()

        if tokens == None:  # 如果收到 None，表示退出
            break

        # 计算 token 序列的 logprobs
        logprob = compute_logprob_of_token_sequence(tokens, context_len, device=device)

        # 将结果放入主进程的队列
        main_queue.put((logprob, shard_id, is_canonical))

def main(dataset_path, context_len=2048, num_shards=5, permutations_per_shard=25,
         random_seed=0, log_file_path=None, max_examples=5000):

    # 设置随机种子，保证可重复性
    random.seed(random_seed)
    np.random.seed(random_seed)

    # 加载数据集
    examples = load_dataset(dataset_path)
    examples = examples[:max_examples]  # 限制加载的示例数量
    num_examples = len(examples)  # 获取数据集大小
    print(f"Loaded {num_examples} examples from {dataset_path}")

    # 对示例进行简单的基于空格的分词
    tokenized_examples = [ex.split() for ex in examples]

    # 使用多进程处理请求（在本例中仅使用一个工作进程）
    processes = []
    main_queue = Queue()  # 主进程队列，用于收集工作进程的结果
    worker_queues = [Queue() for _ in range(1)]  # 工作进程队列

    # 启动工作进程
    p = Process(target=worker, args=(context_len, 0, main_queue, worker_queues[0]))
    processes.append(p)
    p.start()

    # 计算每个分片的大小（将数据集分为多个分片）
    shard_counts = [(x + 1 if i < num_examples % num_shards else x) 
                    for i, x in enumerate([num_examples // num_shards] * num_shards)]
    shard_counts = np.asarray(shard_counts)

    # 生成每个分片的索引
    shard_example_indices = [0] + np.cumsum(shard_counts).tolist()
    for i, (start, end) in enumerate(zip(shard_example_indices, shard_example_indices[1:])):
        shard = tokenized_examples[start:end]

        # 将原始顺序的logprobs请求提交到worker队列
        worker_queues[0].put((
            flatten(shard),  # 展平后的token列表
            i,               # 分片ID
            True))           # 标识这是canonical（原始顺序）

        # 将打乱顺序的logprobs请求提交到worker队列
        for j in range(permutations_per_shard):
            worker_queues[0].put((
                flatten(shuffle(shard)),  # 打乱后的token列表
                i,                        # 分片ID
                False))                   # 标识这是打乱顺序

    # 等待所有请求完成，并显示进度条
    total_work = num_shards * (1 + permutations_per_shard)
    pbar = tqdm(total=total_work)

    canonical_logprobs = [None for _ in range(num_shards)]  # 存储每个分片的 canonical logprobs
    shuffled_logprobs  = [[] for _ in range(num_shards)]    # 存储每个分片的打乱顺序 logprobs

    # 处理worker进程返回的结果
    completed = 0
    while completed < total_work:
        logprob, shard_id, is_canonical = main_queue.get()

        if is_canonical:
            canonical_logprobs[shard_id] = logprob  # 存储原始顺序的logprobs
        else:
            shuffled_logprobs[shard_id].append(logprob)  # 存储打乱顺序的logprobs

        pbar.update(1)  # 更新进度条
        completed += 1

    # 终止工作进程
    worker_queues[0].put((None, None, None))  # 向worker发送退出信号

    for p in processes:
        p.join()  # 等待所有worker进程结束

    # 计算 p-value（p值，用于统计显著性检验）
    canonical_logprobs = np.asarray(canonical_logprobs)  # 转换为numpy数组
    shuffled_logprobs  = np.asarray(shuffled_logprobs)

    # 进行 t 检验，计算 canonical 和 shuffled 之间的差异
    diffs = canonical_logprobs - shuffled_logprobs.mean(axis=1)
    z = np.mean(diffs) / np.std(diffs) * np.sqrt(len(diffs))
    pval = 1 - tdist.cdf(z, df=len(diffs)-1)  # 计算 p 值
    print(f"{pval=}")

    # 将结果写入日志文件（如果指定了log_file_path）
    if log_file_path is not None:
        print(f"Writing logprobs to: {log_file_path}")
        with open(f"{log_file_path}", 'w') as f:
            f.write(json.dumps({
                'pval': pval,
                'permutations_per_shard': permutations_per_shard,
                'num_shards': num_shards,
                'canonical_logprobs': canonical_logprobs.tolist(),
                'shuffled_logprobs': shuffled_logprobs.tolist(),
            }))

if __name__ == '__main__':
  # 使用Fire库，将命令行参数解析并传递给main函数
  fire.Fire(main)

4 结果

{“pval”: 0.1696232809691942, “permutations_per_shard”: 25, “num_shards”: 5, “canonical_logprobs”: [-3.3727318899495287, -26.976947896884596, -8.280387231770165, -11.1112375389544, -15.317197114102502], “shuffled_logprobs”: [[-30.909075783811705, -1.9435003494767502, -6.068986559756483, -26.58900908523132, -16.12269960305306, -4.46379730144066, -2.1558121800502787, -2.7554792693991, -31.682284527334303, -2.6273379016797502, -29.87468835264795, -29.607210920316206, -21.57213257741471, -11.95329938606544, -9.972366131973049, -1.09951527892729, -24.01362313224146, -12.456106343552321, -13.67304127957505, -8.3861853631837, -1.19935666177955, -1.3937543557773802, -1.8002136455626179, -18.009020852617073, -17.578153150829802], [-1.048417377427077, -2.3941244789539704, -8.412327189846044, -24.544644694362702, -5.74892321528065, -1.055017764263738, -9.581590557731854, -7.1433768327109, -2.737799236512142, -23.983025399790144, -9.26424030310054, -14.993957307794997, -2.9504655498746724, -12.080805583617936, -30.364195487487766, -1.9539559864239302, -9.9784173216152, -5.28962663901724, -15.477334895809188, -12.511526170812552, -2.6651197975443917, -7.0888789550949, -3.2381118496705326, -9.586995443264478, -19.668003974102017], [-4.233909482393801, -7.066849438078104, -5.8159291705155995, -10.790016564647766, -5.962899019632103, -5.1748830459693185, -2.6900913199189995, -14.64220487293797, -12.072412084641194, -7.0405692357728995, -3.757379365485161, -4.3277333949891, -18.239703727872094, -1.2460728438048796, -1.5030126277381202, -3.4466863958886114, -4.680143685284249, -4.651795277712018, -20.354000748485447, -1.4471048784444498, -10.138775905959701, -18.178129422154928, -8.530598762226427, -7.489915270562131, -3.1585280028892795], [-11.69612743268735, -7.769248518398181, -6.86862614461919, -6.518956643516701, -3.803943939116793, -13.014972477420848, -8.689137998628949, -15.809698635575, -7.99394916393605, -10.31342520238305, -15.928287922934501, -4.502634127164701, -8.7768807739485, -4.220711983509, -28.029167855395826, -2.81686953962755, -9.31479084741635, -11.158157243828, -6.721924684535599, -10.066673909472277, -4.500597344717, -21.480636945940205, -9.300124195747498, -11.9573958015312, -14.577081120902347], [-6.629690439094301, -10022.289585636432, -7.999280733182201, -11.308060008804496, -5.2692347206537, -7.0436708680337015, -4.26530791126701, -4.130906993623899, -5.9630648153422, -7.950204238607298, -7.942466538914552, -12.449199286857947, -25.78205265161044, -17.262547473382632, -5.530209510927001, -8.1570511425078, -5.8230390775959995, -5.532957394563099, -16.9681575425189, -5.454541042652198, -4.4566292699186, -2.14531132561838, -38.43645063084328, -33.65827228043184, -2.714607539457402]]}