examples/api_eval/eval_openai.py

#!/usr/bin/python3
# -*- coding: utf-8 -*-
import argparse
import json
import os
import sys
import time

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():
    """
    python3 eval_openai.py --model_name gpt-4o --eval_result eval_math_result_gpt-4o.jsonl
    python3 eval_openai.py --model_name gpt-4o-mini --eval_result eval_math_result_gpt-4o-mini.jsonl
    :return:
    """
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--api_key",
        default=environment.get(key="OPENAI_API_KEY"),
        type=str
    )
    parser.add_argument(
        "--model_name",
        default="gpt-4o",
        # default="gpt-4o-mini",
        type=str
    )
    parser.add_argument(
        "--eval_data",
        default=(project_path / "data/arc-easy.jsonl").as_posix(),
        type=str
    )
    parser.add_argument(
        "--eval_result",
        default=(project_path / "data/eval_math_result.jsonl").as_posix(),
        type=str
    )
    args = parser.parse_args()
    return args


def main():
    args = get_args()

    # gpt-4o: 82
    # gemini: 89
    client = AzureOpenAI(
        # api_key=args.api_key,
        api_key="Dqt75blRABmhgrwhfcupd1rq44YqNuEgku8FcFFDrEljMq6gltf0JQQJ99BCACYeBjFXJ3w3AAABACOG2njW",
        api_version="2025-01-01-preview",
        azure_endpoint="https://west-us-chatgpt.openai.azure.com"
    )

    total = 0
    total_correct = 0

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

    with open(args.eval_data, "r", encoding="utf-8") as fin, open(args.eval_result, "a+", encoding="utf-8") as fout:
        for row in fin:
            if total > 20:
                break

            row = json.loads(row)
            idx = row["id"]
            question = row["question"]
            choices = row["choices"]
            answer_key = row["answerkey"]

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

            instruct = "Complete this single-choice question."

            choices_str = ""
            for choice in choices:
                label = choice["label"]
                text = choice["text"]
                choices_str += f"If you think the answer is `{text}` output: `{label}`\n"

            prompt = f"""
{instruct}

Question: 
{question}

Choices: 
{choices_str}

Remember to output ONLY the corresponding letter.
Your output is:
            """.strip()
            # print(prompt)
            try:
                time_begin = time.time()
                response = client.chat.completions.create(
                    model=args.model_name,
                    messages=[{"role": "user", "content": prompt}],
                    stream=False,
                    # max_tokens=1,
                    temperature=0.0,
                    # logit_bias={
                    #     32: 100,
                    #     33: 100,
                    #     34: 100,
                    #     35: 100,
                    #     36: 100,
                    # }
                )
                time_cost = time.time() - time_begin
                print(time_cost)
            except openai.BadRequestError as e:
                print(f"request failed, error type: {type(e)}, error text: {str(e)}")
                continue

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

            correct = 1 if prediction == answer_key else 0

            total += 1
            total_correct += correct
            score = total_correct / total

            row_ = {
                "id": idx,
                "question": question,
                "choices": choices,
                "ground_true": answer_key,
                "prediction": prediction,
                "correct": correct,
                "total": total,
                "total_correct": total_correct,
                "score": score,
                "time_cost": time_cost,
            }
            row_ = json.dumps(row_, ensure_ascii=False)
            fout.write(f"{row_}\n")

            # print(f"score: {score}")

    return


if __name__ == "__main__":
    main()