script/overfit_hardcode.py

import json
import re
from colorama import Fore, init
import os

init(autoreset=True)

# ================= 配置区 =================
INPUT_FILE = "/root/test/weitiao/data_process_bq/data3/sharegpt_h1_nam_num_nhh_len_hard_rej_think_nobitewis_valid_left5.json"
# 建议输出为新的 json 文件，以免覆盖原文件
OUTPUT_FILE = "/root/test/weitiao/data_process_bq/data3/sharegpt_h1_nam_num_nhh_len_hard_rej_think_nobitewis_valid_left5_hcoded.json"
# ==========================================

# ================= 匹配模式编译区 =================
# 将 Prompt 中的描述转为正则表达式 (忽略大小写)
# 针对 "Can I..." 这种词，增加了对省略号(…或...)的匹配，防止将正常的 "Can I help you" 误杀
CLICHE_PATTERNS = [
    # 1. 空壳式互动启动语
    r"\bcan i\s*(?:\.{3,}|…)", 
    r"\bmay i\s*(?:\.{3,}|…)", 
    r"\blet me\s*(?:\.{3,}|…)", 
    r"\byou know\s*(?:\.{3,}|…)", 
    r"\bguess what\s*(?:\.{3,}|…)", 
    r"\bsomething personal\s*\?", 
    r"\bsomething important\s*\?", 
    r"\byou know what i\s*(?:—|-{1,2})", 
    r"\bi have something\s*(?:\.{3,}|…)",
    r"\bi wanted to tell you something",
    r"\bthere['’]s something i should say",
    r"\bi was wondering if",
    # 0309_v3 & 0209_v2 & v1 补充短语
    r"\bdo you want\b",
    r"\bpromise me something\b",
    r"\bcan you\b",
    r"\bwould you\b",
    r"\bcould i steal a private question from you\b",
    r"\bdo you have a secret corner of your time just for me\b",
    r"\bi have a thought that only your ears should hear\b",
    r"\bcan i\b",
    r"\byou know something\??",
    r"\*honest\?",
    r"\*personal\?",
    r"\bsomething important\b",
    r"\bsomething honest\b",
    r"\bsomething personal\b",
    r"\byou know what\b",
    r"\byou know what, sweetheart\?",
    r"\byou know what's crazy\?",
    r"\byou know what's funny\?",
    r"\byou know what i think\?",
    r"\bsomething kind of weird\?",
    r"\byou know what i love\?",
    r"\byou know what that means, right\?",
    r"\byou know what this means, right\?",
    r"\bask something\b",
    r"\bask you something\b",
    r"\bcan i ask\b",
    r"\bi ask you\b",
    r"\bi ask something\b",
    r"\bmay i\b",
    r"\bmay i ask\b",
    r"\bwanna know\b",
    r"\bwanna see\b",
    r"\byou know what\?",
    r"\bdo you know something\b",
    r"\bdo you wanna\b",

    # 2. 咬唇综合征 (兼容 bites/biting, his/her/your, lower lip)
    r"\bbit(e|es|ing) (his|her|my|your)( lower)? lip",

    # 3. 气音与低语狂热
    r"\bvoice barely above a whisper\b", 
    r"\bvoice drops to a whisper\b", 
    r"\bdropping to a sultry\b",
    r"\bdropping to a husky\b",

    # 4. 刻板仰视
    r"\blook(s|ing)? up at (you|him|her|me)\b",

    # 5. 陈腔滥调的生理反应
    r"\bheart skips a beat\b", 
    r"\btake(s|ing)? a deep breath\b", 
    r"\beyes widen(ed)? in shock\b", 
    r"\btears prick at the corners\b",

    # 6. 标志性动作复读
    r"\brun(ning|s)? a hand through (his|her|my|your) hair\b", 
    r"\bmischievous glint in (his|her|my|your) eye(s)?\b",

    # 7. 凑字数模板
    r"\bwith a mix of\b", 
    r"\bfor a moment before\b", 
    r"\bjust like that\b"
]

# 将所有正则组合成一个超级正则，匹配速度极快
COMBINED_PATTERN = re.compile("|".join(CLICHE_PATTERNS), re.IGNORECASE)

def has_cliche(text):
    """检测文本中是否包含上述任意套路短语"""
    if not text:
        return False
    return bool(COMBINED_PATTERN.search(text))

def calculate_dual_quality(chosen_has_cliche, rejected_has_cliche):
    """判断标签的四大规则"""
    if chosen_has_cliche is False and rejected_has_cliche is True:
        return "Perfect"
    elif chosen_has_cliche is False and rejected_has_cliche is False:
        return "Neutral_Clean"
    elif chosen_has_cliche is True and rejected_has_cliche is True:
        return "Bad_Pair"
    elif chosen_has_cliche is True and rejected_has_cliche is False:
        return "Toxic_Reverse"
    else:
        return "Unknown"

def process():
    if not os.path.exists(INPUT_FILE):
        print(f"{Fore.RED}错误：找不到输入文件 {INPUT_FILE}")
        return
    
    print(f"{Fore.CYAN}🚀 开始加载数据...")
    with open(INPUT_FILE, "r", encoding="utf-8") as f:
        data = json.load(f)

    print(f"{Fore.CYAN}🚀 开始进行硬编码规则打标，共计 {len(data)} 条数据...")
    
    # 用于统计四大标签的数量
    stats = {
        "Perfect": 0,
        "Neutral_Clean": 0,
        "Bad_Pair": 0,
        "Toxic_Reverse": 0
    }

    for i, item in enumerate(data):
        # 提取文本
        chosen_text = item.get("chosen", {}).get("value", "")
        rejected_text = item.get("rejected", {}).get("value", "")

        # 正则匹配
        c_c = has_cliche(chosen_text)
        r_c = has_cliche(rejected_text)
        
        # 计算评级
        quality = calculate_dual_quality(c_c, r_c)
        stats[quality] += 1
        
        # 将结果追加到原 json 对象中
        item["audit_result"] = {
            "chosen_has_cliche": c_c,
            "rejected_has_cliche": r_c,
            "dual_quality": quality
        }

        # 终端进度打印
        if i % 500 == 0 and i > 0:
            print(f"[{i}/{len(data)}] 条已处理...")

    print(f"{Fore.CYAN}💾 正在保存结果到: {OUTPUT_FILE} ...")
    with open(OUTPUT_FILE, "w", encoding="utf-8") as f:
        json.dump(data, f, ensure_ascii=False, indent=2)

    # 打印最终统计信息
    print(f"\n{Fore.GREEN}✅ 处理完成！统计结果如下：")
    print(f"{Fore.GREEN}  - Perfect       (Chosen 干净, Rejected 套路): {stats['Perfect']}")
    print(f"{Fore.YELLOW}  - Neutral_Clean (Chosen 干净, Rejected 干净): {stats['Neutral_Clean']}")
    print(f"{Fore.RED}  - Bad_Pair      (Chosen 套路, Rejected 套路): {stats['Bad_Pair']}")
    print(f"{Fore.RED}  - Toxic_Reverse (Chosen 套路, Rejected 干净): {stats['Toxic_Reverse']}")

if __name__ == "__main__":
    process()