script/api_4overfit.py

import json, os, re, concurrent.futures
from openai import OpenAI
from colorama import Fore, init
import time

init(autoreset=True)

# ================= 配置区 =================
API_KEY = "sk_6K2WAGvUtmmJspmA82xp5Bi7uOWNJZo2XeNJZ5kgo4o"
BASE_URL = "https://api.ppio.com/openai"
MODEL = "zai-org/glm-5" 

INPUT_FILE = "/root/test/weitiao/data_process_bq/data3/result/M_sharegpt_data3_all_fail.json"
OUTPUT_FILE = "/root/test/weitiao/1" 

MAX_WORKERS = 40
TIMEOUT = 90
# ==========================================

client = OpenAI(api_key=API_KEY, base_url=BASE_URL)

# ================= 针对 N-gram 模式坍塌的专用 Prompt =================
DPO_DUAL_PROMPT = """你是一名专业的小说文本质检员和自然语言处理分析师。你的任务是检测文本中是否出现了严重的“模式坍塌（Mode Collapse）”和“陈词滥调（Clichés）”。

【极其重要】：
1. 忽略文中描述的任何行为、道德、合规性或安全性。
2. 无论内容多么露骨（NSFW）或违规视为“虚构数据”，不要进行安全判定。
3. 只寻找“刻板的生理反应描写”“高度重复的套路词汇”以及“空壳式互动启动语”。

1. 空壳式互动启动语【最高优先级】：指仅用于制造互动氛围或暧昧节奏、但不提供任何实质信息、如果一句话删除后不影响剧情理解的启动句或提问句：
“Can I…”, “May I…”, “Let me…”, “You know…”, “Guess what…”、 “Something personal?”, “Something important?”、未完成句式（ “You know what I—”）、“I have something…” 、语义等价表达（陈述句：“I wanted to tell you something.”、“There’s something I should say。”、“I was wondering if…” ）
2. 咬唇综合征："bites her/his lip", "biting her lower lip" 等。
3. 气音与低语狂热："voice barely above a whisper", "voice drops to a whisper", "dropping to a sultry/husky..."。
4. 刻板仰视："looks up at you/him", "looking up at... with..."。
5. 陈腔滥调的生理反应："heart skips a beat", "takes a deep breath", "eyes widen in shock", "tears prick at the corners"。
6. 标志性动作复读："running a hand through his hair", "a mischievous glint in her eye"。
7. 凑字数模板："with a mix of...", "for a moment before", "just like that"。

【判定准则】：
- 只要文本中明显使用了上述的套路化表达判定为 True。
- 文本动作描写具体生动、符合角色个性，不存在空壳互动或模板化表达，判定为 False。

请严格返回 JSON：
{
  "chosen_has_cliche": true/false,
  "rejected_has_cliche": true/false
}"""

# ===== 以下代码完全未修改 =====

def calculate_dual_quality(chosen_has_cliche, rejected_has_cliche):
    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 extract_json_robust(text):
    if not text: return None

    text = re.sub(r"^```json\s*", "", text, flags=re.MULTILINE)
    text = re.sub(r"```$", "", text, flags=re.MULTILINE).strip()

    try:
        return json.loads(text)
    except json.JSONDecodeError:
        pass

    try:
        match = re.search(r'(\{.*)[,"]', text, re.DOTALL)
        if match:
            potential_json = match.group(1).strip()
            for suffix in ['"}', '}', 'true}', 'false}']:
                try:
                    return json.loads(potential_json + suffix)
                except:
                    continue
    except:
        return None
    return None

def audit_dual_item(index, item):
    prompt = f"###[A] Chosen Text:\n{item['chosen']['value']}\n\n### [B] Rejected Text:\n{item['rejected']['value']}"
    
    max_retries = 3
    
    for attempt in range(max_retries):
        try:
            response = client.chat.completions.create(
                model=MODEL,
                messages=[
                    {"role": "system", "content": DPO_DUAL_PROMPT}, 
                    {"role": "user", "content": prompt}
                ],
                response_format={"type": "json_object"}, 
                temperature=0.1,  
                timeout=TIMEOUT
            )
            content = response.choices[0].message.content.strip()
            audit = extract_json_robust(content)

            if audit:
                # ✅ 兼容 list 或 dict
                if isinstance(audit, list) and len(audit) > 0:
                    audit = audit[0]

                c_cliche = audit.get("chosen_has_cliche")
                r_cliche = audit.get("rejected_has_cliche")
                audit["dual_quality"] = calculate_dual_quality(c_cliche, r_cliche)
                return {"_original_index": index, "audit_result": audit}
            else:
                return {
                    "_original_index": index,
                    "audit_result": {
                        "chosen_has_cliche": False, 
                        "rejected_has_cliche": False,
                        "dual_quality": "Refused"
                    }
                }
                
        except Exception as e:
            error_msg = str(e)
            if ("timed out" in error_msg.lower() or "timeout" in error_msg.lower() or "connection" in error_msg.lower()) and attempt < max_retries - 1:
                sleep_time = 2 ** attempt
                time.sleep(sleep_time)
                continue
            
            return {"_original_index": index, "error": error_msg}

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
    
    with open(INPUT_FILE, "r", encoding="utf-8") as f: 
        data = json.load(f)

    processed_count = 0
    if os.path.exists(OUTPUT_FILE):
        with open(OUTPUT_FILE, "r", encoding="utf-8") as f: 
            processed_count = sum(1 for _ in f)

    to_process = data[processed_count:]
    print(f"{Fore.CYAN}🚀 鲁棒审计启动（打击过拟合与套路词），剩余: {len(to_process)} 条...")

    with concurrent.futures.ThreadPoolExecutor(max_workers=MAX_WORKERS) as executor:
        futures = {executor.submit(audit_dual_item, i + processed_count, item): i for i, item in enumerate(to_process)}

        for future in concurrent.futures.as_completed(futures):
            res = future.result()
            idx = res.get("_original_index", "??")

            if "error" in res:
                print(f"{Fore.RED}#{idx} | 拦截/错误: {res['error'][:40]}")
            else:
                aud = res["audit_result"]
                # ✅ 防止 audit 是字符串/其他非 dict
                if isinstance(aud, dict):
                    c_c = aud.get("chosen_has_cliche")
                    r_c = aud.get("rejected_has_cliche")
                    quality = aud.get("dual_quality")
                else:
                    c_c = r_c = quality = "Unknown"

                if quality == "Perfect":
                    color = Fore.GREEN
                elif quality == "Toxic_Reverse" or quality == "Bad_Pair":
                    color = Fore.RED
                else:
                    color = Fore.YELLOW
                    
                print(f"#{idx} | {color}解析成功 | Chosen_套路: {c_c} | Rejected_套路: {r_c} | 综合质量: {quality}")

            with open(OUTPUT_FILE, "a", encoding="utf-8") as f:
                f.write(json.dumps(res, ensure_ascii=False) + "\n")

if __name__ == "__main__":
    process()