#!/usr/bin/env python3 # -*- coding: utf-8 -*- import os import json import argparse import logging import threading from tqdm import tqdm from concurrent.futures import ThreadPoolExecutor, as_completed import anthropic logger = logging.getLogger(__name__) logging.basicConfig(level=logging.INFO) # 抑制 httpx / urllib3 的 INFO 日志，给 tqdm 留干净的 stderr logging.getLogger("httpx").setLevel(logging.WARNING) logging.getLogger("urllib3").setLevel(logging.WARNING) # 线程本地存储，用来缓存每个线程的模型实例 thread_local = threading.local() # 用于多线程写文件的互斥锁 write_lock = threading.Lock() def check_answer(response, answer): """ 从 response 的末尾向前查找匹配正确答案。 如果匹配到，返回是否正确；如果未匹配到，随机返回一个选项。 参数: response (str): 模型的输出内容。 answer (str): 正确答案，是一个从 'a' 到 'j' 的小写字母。 返回: bool: 如果预测与正确答案匹配，返回 True；否则返回 False。 """ # 定义所有可能的选项 all_choices = [chr(i) for i in range(ord('a'), ord('j') + 1)] if response == None: return False, random.choice(all_choices) # 清理 response 的尾部标点符号 for char in [",", ".", "!", "?", ";", ":", "'"]: response = response.strip(char) response = " " + response.lower() + " " # 添加空格并将响应转换为小写，以避免大小写问题 import re # 使用正则表达式提取 : 后面的第一个字母 match = re.search(r': (\w)', response) # \w 匹配字母或数字，这里主要提取字母 if match: extracted_answer = match.group(1) # 提取第一个字符 # 比较提取的答案与正确答案 if extracted_answer == answer: return True, extracted_answer # 如果匹配正确，返回 True 和答案 return False, extracted_answer # 如果匹配错误，返回 False 和提取的答案 # 初始化候选选项 candidates = [] # 按从后往前的规则查找 for choice in reversed(all_choices): # 检查括号形式 (a), (b), ... if f"({choice})" in response: candidates.append(choice) continue # 检查空格分隔形式 a, b, ... if f" {choice} " in response: candidates.append(choice) continue # 检查带点形式 a., b., ... if f"{choice}." in response: candidates.append(choice) # 检查前后不是英文字母的情况 choice_pos = response.find(choice) if choice_pos != -1: prev_char = response[choice_pos - 1] if choice_pos > 0 else '' next_char = response[choice_pos + 1] if choice_pos + 1 < len(response) else '' if prev_char not in string.ascii_lowercase and prev_char not in string.ascii_uppercase and \ next_char not in string.ascii_lowercase and next_char not in string.ascii_uppercase: candidates.append(choice) # 如果找到多个候选项，选择最后一个匹配到的 if len(candidates) > 0: pred_index = candidates[-1] # 最后匹配到的候选项 else: # 如果没有匹配到候选项，随机选择 pred_index = random.choice(all_choices) # 返回预测结果是否等于正确答案 return pred_index == answer[0], pred_index def process_record(rec, args): """ 对单条记录： 1) 拼 image_parts + text 2) 调用模型 3) 判断是否正确 4) 返回结果 dict """ vk = rec["video_key"] messages = rec["messages"] client = anthropic.Anthropic( base_url=args.base_url, api_key=args.api_key ) try: response = client.messages.create( model=args.model_name, messages=messages, max_tokens=args.max_tokens, timeout=120 # 秒 ) prediction = response.choices[0].message.content.strip() except Exception as e: logger.error(f"[{vk}] 调用模型失败: {e}") response = "" prediction = "" correct = rec.get("answer", "") is_correct = check_answer(prediction,correct) result = { "video_key": vk, "question_id": rec["question_id"], "model_response": prediction, "correct": correct, "is_true": is_correct, } if rec.get("question_type"): result["question_type"] = rec["question_type"] return result def save_result(output_file, record): """ 多线程安全地 append 写入 JSONL，并打印 debug 回执 """ try: with write_lock: with open(output_file, "a", encoding="utf-8") as f: f.write(json.dumps(record, ensure_ascii=False) + "\n") f.flush() logger.debug(f"写入成功：{record['video_key']}/{record['question_id']}") except Exception as e: logger.error(f"写入失败 [{output_file}]：{e}，record={record}") def main(): parser = argparse.ArgumentParser() parser.add_argument("--messages_file", required=True, help="输入的 JSONL 文件，包含 video_key, question_id, messages, answer 等字段") parser.add_argument("--output_file", required=True, help="推理结果输出 JSONL，每行一个结果") parser.add_argument("--api_key", required=True, help="调用模型的 API key") parser.add_argument("--base_url", required=True, help="调用模型的 API key") parser.add_argument("--model_name", default="gemini-2.0-flash", help="模型名称") parser.add_argument("--max_tokens", type=int, default=100, help="生成最大 token 数") parser.add_argument("--workers", type=int, default=4, help="并发线程数") parser.add_argument("--no-save", action="store_true", help="不写入结果文件") parser.add_argument("--no-progress", action="store_true", help="禁用所有 tqdm 进度条") args = parser.parse_args() # 确认你在看的是正确的输出文件 abs_out = os.path.abspath(args.output_file) logger.info(f"结果文件路径：{abs_out}") # 先清空旧文件 if not args.no_save: with open(args.output_file, "w", encoding="utf-8"): pass logger.info("已清空旧文件，准备写入新结果") # —— 第一步：加载并收集每个 video_key 的 image_parts —— # records = [] # 统计总行数，用于进度条 with open(args.messages_file, "r", encoding="utf-8") as f: total_lines = sum(1 for _ in f) with open(args.messages_file, "r", encoding="utf-8") as f: for line in tqdm(f, total=total_lines, desc="加载 messages_file", disable=args.no_progress): rec = json.loads(line) records.append(rec) # —— 第二步：并发推理 —— # with ThreadPoolExecutor(max_workers=args.workers) as exe: futures = { exe.submit(process_record, rec, args): rec for rec in records } iterator = as_completed(futures) if not args.no_progress: iterator = tqdm(iterator, total=len(futures), desc="推理中", disable=args.no_progress) for fut in iterator: res = fut.result() if not args.no_save: save_result(args.output_file, res) # —— 完成后，统计写入行数 —— # if not args.no_save: try: cnt = sum(1 for _ in open(args.output_file, "r", encoding="utf-8")) logger.info(f"总共写入 {cnt} 行结果") except Exception: pass logger.info("全部推理完成") if __name__ == "__main__": main()