llm_eval_script/azure_openai_summary.py

#!/usr/bin/python3
# -*- coding: utf-8 -*-
import argparse
import re
from datetime import datetime
import json
import os
from pathlib import Path
import sys
import time
from zoneinfo import ZoneInfo  # Python 3.9+ 自带，无需安装

pwd = os.path.abspath(os.path.dirname(__file__))
sys.path.append(os.path.join(pwd, "../"))

import openai
from openai import AzureOpenAI

from project_settings import environment, project_path


def get_args():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--model_name",
        # default="gpt-4o",
        default="gpt-4o-mini",
        type=str
    )
    parser.add_argument(
        "--eval_dataset_name",
        default="agent-bingoplus-ph-25-summary.jsonl",
        type=str
    )
    parser.add_argument(
        "--eval_dataset_dir",
        default=(project_path / "data/dataset").as_posix(),
        type=str
    )
    parser.add_argument(
        "--eval_data_dir",
        default=(project_path / "data/eval_data").as_posix(),
        type=str
    )
    parser.add_argument(
        "--client",
        default="shenzhen_sase",
        type=str
    )
    parser.add_argument(
        "--service",
        default="west_us_chatgpt_openai_azure_com",
        type=str
    )
    parser.add_argument(
        "--create_time_str",
        default="null",
        # default="20250806_114802",
        type=str
    )
    parser.add_argument(
        "--interval",
        default=5,
        type=int
    )
    args = parser.parse_args()
    return args


def main():
    args = get_args()

    eval_dataset_dir = Path(args.eval_dataset_dir)
    eval_dataset_dir.mkdir(parents=True, exist_ok=True)
    eval_data_dir = Path(args.eval_data_dir)
    eval_data_dir.mkdir(parents=True, exist_ok=True)

    if args.create_time_str == "null":
        tz = ZoneInfo("Asia/Shanghai")
        now = datetime.now(tz)
        create_time_str = now.strftime("%Y%m%d_%H%M%S")
        # create_time_str = "20250729-interval-5"
    else:
        create_time_str = args.create_time_str

    eval_dataset = eval_dataset_dir / args.eval_dataset_name

    output_file = eval_data_dir / f"azure_openai/azure/{args.model_name}/{args.client}/{args.service}/{create_time_str}/{args.eval_dataset_name}"
    output_file.parent.mkdir(parents=True, exist_ok=True)

    service_params = environment.get(args.service, dtype=json.loads)
    client = AzureOpenAI(
        **service_params,
        # api_key="Dqt75blRABmhgrwhfcupd1rq44YqNuEgku8FcFFDrEljMq6gltf0JQQJ99BCACYeBjFXJ3w3AAABACOG2njW",
        # api_version="2025-01-01-preview",
        # azure_endpoint="https://west-us-chatgpt.openai.azure.com"
    )

    total = 0
    total_score = 0

    # finished
    finished_idx_set = set()
    if os.path.exists(output_file.as_posix()):
        with open(output_file.as_posix(), "r", encoding="utf-8") as f:
            for row in f:
                row = json.loads(row)
                idx = row["idx"]
                total = row["total"]
                total_score = row["total_score"]
                finished_idx_set.add(idx)
    print(f"finished count: {len(finished_idx_set)}")

    with open(eval_dataset.as_posix(), "r", encoding="utf-8") as fin, open(output_file.as_posix(), "a+", encoding="utf-8") as fout:
        for row in fin:
            row = json.loads(row)
            idx = row["idx"]
            system_prompt: str = row["system_prompt"]
            user_prompt: str = row["user_prompt"]
            response = row["response"]

            if idx in finished_idx_set:
                continue
            finished_idx_set.add(idx)

            messages = [
                {
                    "role": "system",
                    "content": system_prompt
                },
                {
                    "role": "user",
                    "content": user_prompt
                },
            ]

            try:
                time.sleep(args.interval)
                print(f"sleep: {args.interval}")
                time_begin = time.time()
                llm_response = client.chat.completions.create(
                    model=args.model_name,
                    messages=messages,
                    stream=False,
                    # max_tokens=1,
                    top_p=0.95,
                    temperature=0.6,
                    # logit_bias={
                    #     32: 100,
                    #     33: 100,
                    #     34: 100,
                    #     35: 100,
                    #     36: 100,
                    #     37: 100,
                    # }
                )
                time_cost = time.time() - time_begin
                print(f"time_cost: {time_cost}")
            except openai.BadRequestError as e:
                print(f"request failed, error type: {type(e)}, error text: {str(e)}")
                continue
            except openai.InternalServerError as e:
                print(f"request failed, error type: {type(e)}, error text: {str(e)}")
                continue

            prediction = llm_response.choices[0].message.content

            response_ = json.loads(response)
            response_tag_name_list = response_["tag_name_list"]
            # print(response_tag_name_list)

            if prediction.startswith("```json") and prediction.endswith("```"):
                prediction_ = prediction[7:-3]
            else:
                prediction_ = prediction

            prediction_tag_name_list = list()
            try:
                prediction_ = json.loads(prediction_)
                prediction_tag_name_list = prediction_["tag_name_list"]
            except json.JSONDecodeError:
                pass
            # print(prediction_tag_name_list)

            # recall
            recall_count = 0
            for tag in response_tag_name_list:
                if tag in prediction_tag_name_list:
                    recall_count += 1
            recall = recall_count / (len(response_tag_name_list) + 1e-7)

            # precision
            precision_count = 0
            for tag in prediction_tag_name_list:
                if tag in response_tag_name_list:
                    precision_count += 1
            precision = precision_count / (len(prediction_tag_name_list) + 1e-7)

            # f1
            f1 = 2 * (recall * precision) / (recall + precision + 1e-7)

            total += 1
            total_score += f1
            score = total_score / total

            row_ = {
                "idx": idx,
                "system_prompt": system_prompt,
                "user_prompt": user_prompt,
                "response": response,
                "prediction": prediction,
                "recall": recall,
                "precision": precision,
                "f1": f1,
                "total": total,
                "total_score": total_score,
                "score": score,
                "time_cost": time_cost,
            }
            row_ = json.dumps(row_, ensure_ascii=False)
            fout.write(f"{row_}\n")
            fout.flush()

    return


if __name__ == "__main__":
    main()