llm_as_judge.py

import heapq
import math
import random
import re
import json
from typing import List, Tuple, Dict, Any, Optional
import itertools
from transformers import AutoTokenizer
import asyncio  # New import added for async handling
from openai import AsyncOpenAI   # Using AsyncOpenAI as client
import numpy as np
from openai import OpenAI
import openai
import json
from tqdm import tqdm 
import re
def read_jsonl(file_path):
    data = []
    with open(file_path, 'r', encoding='utf-8') as file:
        for line in file:
            data.append(json.loads(line.strip()))
    return data

def extract_answer_judge(solution_text: str):
    boxed_pattern = r'\\boxed\{([^}]*)\}'
    matches = re.findall(boxed_pattern, solution_text)
    if matches:
        return matches[-1].strip()
    return None

def separate_steps(steps: List[str], mode: str = 'join') -> Any:
    delimiter = "\n\n"
    if mode == 'join':
        if not isinstance(steps, list):
            raise TypeError("For 'join' mode, 'steps' must be a list of strings.")
        return delimiter.join(steps)
    elif mode == 'split':
        if not isinstance(steps, str):
            raise TypeError("For 'split' mode, 'steps' must be a string.")
        return steps.split(delimiter)
    else:
        raise ValueError("Mode should be either 'join' or 'split'.")
    
    
    
def evaluate_llm_as_judge(problem: str, steps: list, output_type: str = 'bool') -> bool:
    global client

    client = OpenAI(
    base_url="http://localhost:8014/v1",
    api_key="token-abc123"
        )
    
    # client = OpenAI(
    # base_url="https://ark.cn-beijing.volces.com/api/v3",
    # api_key="d61217e7-8ff3-4937-83ed-3dd2bebf72ad"
    #     )

    model_name = "DeepSeek-R1-Distill-Qwen-14B"
    # model_name = 'deepseek-v3-241226'
    messages = []
    feedback = None

    judge_prompt = f"""
    The following is a math problem and a solution (split into paragraphs, enclosed with tags and indexed from 0):

    [Math Problem]

    {problem}

    [Solution]

    {steps}

    Your task is to review and critique the solution paragraph by paragraph. Once you identify an error in a paragraph, return the index of the paragraph where the earliest error occurs. Otherwise, return the index of -1 (which typically denotes "not found").

    Please put your final answer (i.e., the index) in \\boxed{{}}.

    """
    messages.append({
            'role': 'user',
            'content': judge_prompt
        })
    completion = client.chat.completions.create(
        model=model_name,
        messages=messages,
        n=1,
        temperature=0.6,
        max_tokens=8192,
    )
    response = completion.choices[0].message.content
    # print('*****step*****',steps)
    # print("*****Verification*****:", response)
        

    content = response.strip().lower()
    last_words = ' '.join(content.split()[-10:])  # Last 3 words

    # print('last_words:', last_words)
    error_step_index = extract_answer_judge(last_words)
    
    if isinstance(error_step_index, int) and (int(error_step_index) != -1) and (error_step_index != None):
        print('error_step_index',error_step_index)
        merged_data = {
            'question': problem,
            'reasining_steps': steps,
            'error_step_index': error_step_index,
            'response': response
        }
        return merged_data
    
    elif int(error_step_index) == -1:
        print('error_step_index',error_step_index)
        merged_data = {
            'question': problem,
            'reasining_steps': steps,
            'error_step_index': -1,
            'response': response
        }
        return merged_data
    
    else:
        return None
    
    
new_file_path = '0312_training_new_processed_16w.jsonl'
data_all = read_jsonl(new_file_path)
print(len(data_all))
output =  []
zero = 0
for data in tqdm(data_all[0:10]):
    print(data_all.index(data))
    problem = data['question']
    steps_ori = data['process']
    labels = data['label']
    steps  = steps_ori.split('\n\n')
    steps[0] = problem + ' ' + steps[0]
    # print('steps:',steps)
    steps_updated = steps[0:len(steps)-1]
    
    
    if zero in data['label']:

        merged_data = evaluate_llm_as_judge(problem=problem, steps=steps_updated, output_type='bool')
        if merged_data != None:
            output.append(merged_data)
        
    else:
        merged_data = {
            'question': problem,
            'reasining_steps': steps_updated,
            'error_step_index': -1,
            'response': '<think>\n\n</think>-1'
        }
        output.append(merged_data)
    

    if len(output) % 100 == 0:
        output_file = '0312_training_fast_slow_thinking.jsonl'
        with open(output_file, 'w', encoding='utf-8') as output_file:
            for entry in output:
                output_file.write(json.dumps(entry, ensure_ascii=False) + '\n')

        print(f"数据已成功写入 {output_file}")