llm_eval_script/siliconflow.py

#!/usr/bin/python3
# -*- coding: utf-8 -*-
"""
https://cloud.siliconflow.cn/sft-d1rosn8o8n4s73ftpa1g/playground/chat/17885302852

https://docs.siliconflow.cn/cn/userguide/capabilities/reasoning

Model Name:
Pro/deepseek-ai/DeepSeek-R1
Tips:
(1)thinking_budget: Must be greater than or equal to 1
(2)The selected model requires paid balance. Your paid balance is insufficient. Please top up and try again.

Model Name:
tencent/Hunyuan-A13B-Instruct
Tips:
(1)它在回答时总是会先思考,最后给出答案.这适合知识问答,但不符合我们Agent的需求. 后来强制其只能输出 A-E 中的一个字符(max_tokens=4),以完成评估.
max_tokens=4,
logit_bias={
    32: 100,
    33: 100,
    34: 100,
    35: 100,
    36: 100,
    37: 100,
},

Model Name:
deepseek-ai/DeepSeek-R1
Tips:
(1)为了让它只输出一个字符,设置 max_tokens=3


Model Name:
Qwen/Qwen3-8B
deepseek-ai/DeepSeek-R1-0528-Qwen3-8B
deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
deepseek-ai/DeepSeek-V3
Tips:
(1)为了让它只输出一个字符,设置 max_tokens=1



Model Name:
baidu/ERNIE-4.5-300B-A47B
Tips:
(1)它可能使用的是bpe 分词, logit_bias 注释掉。

"""
import argparse
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, "../"))

from openai import OpenAI

from project_settings import environment, project_path


def get_args():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--model_name",
        # default="Pro/deepseek-ai/DeepSeek-R1",
        # default="tencent/Hunyuan-A13B-Instruct",
        default="deepseek-ai/DeepSeek-V3",
        # default="Qwen/Qwen3-8B",
        # default="deepseek-ai/DeepSeek-R1",
        # default="deepseek-ai/DeepSeek-R1-0528-Qwen3-8B",
        # default="deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",
        # default="baidu/ERNIE-4.5-300B-A47B",
        type=str
    )
    parser.add_argument(
        "--eval_dataset_name",
        default="agent-bingoplus-ph-90-choice.jsonl",
        # default="agent-lingoace-zh-400-choice.jsonl",
        # default="arc-easy-1000-choice.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="siliconflow_api_key",
        type=str
    )
    parser.add_argument(
        "--create_time_str",
        default="null",
        # default="20250728_113641",
        type=str
    )
    parser.add_argument(
        "--interval",
        default=1,
        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 = "20250724_090615"
    else:
        create_time_str = args.create_time_str

    eval_dataset = eval_dataset_dir / args.eval_dataset_name

    model_name_ = args.model_name.replace("/", "#")
    output_file = eval_data_dir / f"siliconflow/siliconflow/{model_name_}/{args.client}/{args.service}/{create_time_str}/{args.eval_dataset_name}"
    output_file.parent.mkdir(parents=True, exist_ok=True)

    api_key = environment.get(args.service, dtype=str)
    client = OpenAI(
        base_url="https://api.siliconflow.cn/v1/",
        # Read your Ark API Key from the environment variable.
        api_key=api_key
    )

    total = 0
    total_correct = 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_correct = row["total_correct"]
                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"]
            prompt = row["prompt"]
            response = row["response"]

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

            try:
                time.sleep(args.interval)
                print(f"sleep: {args.interval}")
                time_begin = time.time()
                completion = client.chat.completions.create(
                    model=args.model_name,
                    messages=[
                        {"role": "user", "content": prompt},
                    ],
                    stream=False,
                    # max_tokens=4096,
                    max_tokens=1,
                    temperature=0.6,
                    top_p=0.95,
                    logit_bias={
                        32: 100,
                        33: 100,
                        34: 100,
                        35: 100,
                        36: 100,
                        37: 100,
                        38: 100,
                        39: 100,
                    },
                    extra_body={
                        "thinking_budget": 1
                    }
                )
                time_cost = time.time() - time_begin
                print(f"time_cost: {time_cost}")
            except Exception as e:
                print(f"request failed, error type: {type(e)}, error text: {str(e)}")
                continue

            # print(f"completion: {completion}")
            prediction = completion.choices[0].message.content

            correct = 1 if prediction == response else 0

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

            row_ = {
                "idx": idx,
                "prompt": prompt,
                "response": response,
                "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")
            fout.flush()

    return


if __name__ == "__main__":
    main()