Skip to main content

一个面向运营商领域的大模型人机测评工具,支持本地与API模型评估

Project description

📡televal-电信运营商人机测评工具

televal 是一个用于评估大语言模型(LLMs)在电信运营商领域任务中表现的测评工具包,支持本地模型与 API 模型的推理调用,内置客观题与主观题的评估方法,以及题型维度的可视化结果输出。

📦源码安装

pip install televal

📁项目目录结构要求

your_project/
├── datas/                  # 存放待评估数据集(如 A.json / B.json)
├── outputs/                # 自动生成,存储模型推理与评估结果
├── configs/
│   ├── model_configs.py    # 配置本地模型名称和路径
│   └── api_models.py       # 配置 API 模型和用于主观评估的裁判LLM
├── televal/                # 安装后的主程序包(无需手动改动)

📄数据准备

将测评数据集(如 A.json)上传至 your_project/datas/ 目录中。我们已提供示例数据集,可用于功能验证。

🧠模型配置

1. 本地模型(可选)

configs/model_configs.py 中加入模型名称及其路径:

model_configs = {
    'qwen2.5-7b-instruct': '/path/to/Qwen2.5-7B-Instruct',
    'qwen2.5-14b-instruct': '/path/to/Qwen2.5-14B-Instruct',
}

2. API 模型(可选)

configs/api_models.py 中配置两个函数:

from openai import OpenAI

# ✅ 必填:裁判模型(用于主观题评估)
def model_eval(prompt):
    client = OpenAI(api_key="YOUR_KEY", base_url="https://api.example.com")
    chat_completion = client.chat.completions.create(
        messages=[{ "role": "user", "content": prompt }],
        model="DeepSeek-V3",
    )
    return chat_completion.choices[0].message.content

# ✅ 可选:评估模型(用于推理答题),当以 API 模式调用时需填写
def model_call(prompt, model_name):
    client = OpenAI(api_key="YOUR_KEY", base_url="https://api.example.com")
    chat_completion = client.chat.completions.create(
        messages=[{ "role": "user", "content": prompt }],
        model=model_name,
    )
    result = chat_completion.choices[0].message.content
    return result

🚀一键评估使用示例

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "2,3"

from televal.generation import generator_local, generator_api
from televal.evaluation import compute_metrics, compute_metrics_LLM, compute_scores
from televal.visualization import compute_ranking_average
import pandas as pd
import IPython.display as disp

def run_all_steps(dataset_name, model, gpu_ids=None, is_api=False):
    print("\n Step 1: 生成回答")
    if is_api:
        generator_api.main(dataset_name=dataset_name, model=model)
    else:
        generator_local.main(dataset_name=dataset_name, model=model, gpu_ids=gpu_ids)

    print("\n Step 2: 客观题评估")
    compute_metrics.main(dataset_name=dataset_name, model=model)

    print("\n Step 3: 主观题评估")
    compute_metrics_LLM.main(dataset_name=dataset_name, model=model)

    print("\n Step 4: 综合得分")
    compute_scores.main(dataset_name=dataset_name, model=model)

    print("\n Step 5: 可视化输出")
    df_type, df_cat = compute_ranking_average.main(dataset_name=[dataset_name])
    disp.display(df_type)
    disp.display(df_cat)

# 示例1:使用本地模型评估
run_all_steps(dataset_name="A", model="qwen2.5-14b-instruct", gpu_ids=[2, 3], is_api=False)

# 示例2:使用 API 模型评估
run_all_steps(dataset_name="B", model="DeepSeek-V3", is_api=True)

📊输出结果说明

运行后,系统将自动生成以下输出文件至 outputs/{dataset}/{model}/

  • record.json:模型推理记录
  • evaluation.json:记录了主观题的逐题评分详情
  • result.json:整合主观题和客观题评估准确率
  • score.json:题型加权得分
  • 题型排名.csv维度排名.csv:在 outputs/ 下生成的可视化评估结果

📬联系我们

如有需求或建议,欢迎联系:wang.yingying@ustcinfo.com

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

televal-0.1.1.tar.gz (10.0 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

televal-0.1.1-py3-none-any.whl (16.6 kB view details)

Uploaded Python 3

File details

Details for the file televal-0.1.1.tar.gz.

File metadata

  • Download URL: televal-0.1.1.tar.gz
  • Upload date:
  • Size: 10.0 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.1.0 CPython/3.10.0

File hashes

Hashes for televal-0.1.1.tar.gz
Algorithm Hash digest
SHA256 cfcbc0295d526717f722a0505de7f0720b4bc73fdcbc9c0561d5e720197139b4
MD5 a0b475aa8ff925b5c5f44cf470763d97
BLAKE2b-256 6be872b0f6f9c5424776db452038b2149716dfcc384b341dbd36f3a0dfd78e7e

See more details on using hashes here.

File details

Details for the file televal-0.1.1-py3-none-any.whl.

File metadata

  • Download URL: televal-0.1.1-py3-none-any.whl
  • Upload date:
  • Size: 16.6 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.1.0 CPython/3.10.0

File hashes

Hashes for televal-0.1.1-py3-none-any.whl
Algorithm Hash digest
SHA256 8814feff16d42852418be7d615e4e9bdc0d91cb9af772af91482a892e24c5274
MD5 2b1c4e4d181621197fdf453733ee9b0a
BLAKE2b-256 909b8264a5ad80cd46f3e8d88473d8091a6dc566218fcf50462f751bb678c3fb

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page