test_time_scaling/concept_difficulty_augment_extract.py

import os
import json
import re
from glob import glob

# 目标根目录（可根据需要修改）
DATA_ROOT = 'data/concept_difficulty_augment/Qwen__Qwen2.5-7B-Instruct/abstract_algebra/harder'
OUTPUT_ROOT = os.path.join(DATA_ROOT, 'extracted_entries')

# 正则表达式，提取 JSON 格式的 question/options
QUESTION_JSON_RE = re.compile(r'\{\s*"question"\s*:\s*".*?",\s*"options"\s*:\s*\{.*?\}\s*\}', re.DOTALL)

# 递归查找所有 all_results.json
all_results_files = glob(os.path.join(DATA_ROOT, 'all_results.json'))

os.makedirs(OUTPUT_ROOT, exist_ok=True)

for results_path in all_results_files:
    with open(results_path, 'r', encoding='utf-8') as f:
        results = json.load(f)
    for idx, entry in enumerate(results):
        model_output = entry.get('model_output', '')
        # 只提取第一个合法 JSON 片段
        match = QUESTION_JSON_RE.search(model_output)
        if not match:
            continue  # 跳过无法提取的
        try:
            parsed = json.loads(match.group())
        except Exception as e:
            continue  # 跳过解析失败的
        # 保存为 question_XXXX.json
        fname = f'question_{idx:04d}.json'
        with open(os.path.join(OUTPUT_ROOT, fname), 'w', encoding='utf-8') as f:
            json.dump(parsed, f, ensure_ascii=False, indent=2)
print(f'Extracted {len(results)} entries to {OUTPUT_ROOT}/')