vlmtest/shuffle_new.py

import json
import random
import os
import copy
from collections import defaultdict

json_dir = '/home/data2/hmguo/vlmtest/benchmark'
output_dir = '/home/data2/hmguo/vlmtest/benchmark_shuffle_new'
os.makedirs(output_dir, exist_ok=True)

global_answer_count = defaultdict(int)  # 全局答案计数器，控制平衡

def get_best_shuffle(item, global_answer_count):
    choices = [
        ("choice_a", "choice_a_chinese", "A"),
        ("choice_b", "choice_b_chinese", "B"),
        ("choice_c", "choice_c_chinese", "C"),
        ("choice_d", "choice_d_chinese", "D")
    ]
    
    # 尝试多种 shuffle，选择使得答案分布最均衡的一个
    best_shuffle = None
    min_variance = float('inf')
    for _ in range(10):  # 尝试 10 次打乱
        temp_choices = copy.deepcopy(choices)
        random.shuffle(temp_choices)
        for i, (choice_key, _, original_answer) in enumerate(temp_choices):
            if item["answer"] == original_answer:
                simulated_answer = chr(ord('A') + i)
                break
        
        simulated_count = global_answer_count.copy()
        simulated_count[simulated_answer] += 1

        values = list(simulated_count.values())
        variance = max(values) - min(values)  # 简单衡量均匀性（方差也可）
        
        if variance < min_variance:
            min_variance = variance
            best_shuffle = temp_choices
            best_answer = simulated_answer

    return best_shuffle, best_answer

for json_file in os.listdir(json_dir):
    if not json_file.endswith('.json'):
        continue

    json_path = os.path.join(json_dir, json_file)
    with open(json_path, 'r', encoding='utf-8') as file:
        data = json.load(file)

    for item in data:
        best_choices, new_answer = get_best_shuffle(item, global_answer_count)

        new_data = {}
        for i, (choice_key, chinese_key, _) in enumerate(best_choices):
            new_choice_key = f"choice_{chr(ord('a') + i)}"
            new_chinese_key = f"{new_choice_key}_chinese"

            new_data[new_choice_key] = item[choice_key]
            new_data[new_chinese_key] = item[chinese_key]

        new_data["answer"] = new_answer
        global_answer_count[new_answer] += 1
        item.update(new_data)

    output_path = os.path.join(output_dir, json_file)
    with open(output_path, 'w', encoding='utf-8') as file:
        json.dump(data, file, indent=4, ensure_ascii=False)

    print(f"{json_file} 答案分布: {dict(global_answer_count)}")