evalstone/evalscope/docs/zh/benchmarks/job_bench.md

JobBench

概述

JobBench 评估智能体系统在真实专业工作场景中的表现，这些任务要求阅读参考文件、生成交付物，并整合多源信息。本适配器使用 ModelScope 数据集 evalscope/job-bench。

任务描述

• 任务类型：智能体专业工作 / 交付物生成
• 输入：职场风格的任务提示，可选包含 reference_files/ 目录下的参考文件
• 输出：最终交付物文件写入 jobbench_output/ 目录
• 数据集：ModelScope evalscope/job-bench
• 评估指标：加权 LLM 评分标准得分（normalized_score）

评估说明

• 默认评估划分是 main。
• 配置 judge.models 用于评分标准打分。
• normalized_score 是主要的加权得分（原始总分 / 满分）；pass_rate 是完全通过评分项的未加权比例；judge_score 是通过评分项权重的原始总和，仅作为诊断指标报告。
• Docker 运行默认使用 python:3.11-slim-bookworm 镜像。正式评估时，请提供包含任务所需 Office、PDF 和电子表格工具的镜像。

属性

属性	值
基准测试名称	job_bench
数据集ID	evalscope/job-bench
论文	无
标签	Agent, Knowledge, MultiTurn
指标	normalized_score, pass_rate, judge_score
默认示例数	0-shot
评估划分	main

数据统计

指标	值
总样本数	65
提示词长度（平均）	3344.32 字符
提示词长度（最小/最大）	2000 / 4920 字符

样例示例

子集: default

{
  "input": [
    {
      "id": "9cdb8ee0",
      "content": "You are preparing the Statistical Analysis Plan for Phase III trial XR-2847 evaluating cardiovascular disease prevention. The sponsor requires validation of statistical assumptions against empirical data and regulatory alignment before protoc ... [TRUNCATED 2026 chars] ... erables.\n\nWrite every final deliverable file under `jobbench_output`. Do not put intermediate scratch files there.\nYour final message may summarize what you produced, but files requested by the task must be actual files in\n`jobbench_output`.\n"
    }
  ],
  "id": 0,
  "group_id": 0,
  "tools": [
    {
      "name": "bash",
      "description": "Execute a bash command inside the sandbox environment. Returns the combined stdout / stderr output of the command.",
      "parameters": {
        "properties": {
          "command": {
            "type": "string",
            "description": "The bash command to execute."
          },
          "timeout": {
            "type": "number",
            "description": "Maximum execution time in seconds (default: 60).",
            "default": 60
          }
        },
        "required": [
          "command"
        ]
      }
    },
    {
      "name": "python_exec",
      "description": "Execute Python source code inside the sandbox environment. Returns stdout and stderr output.",
      "parameters": {
        "properties": {
          "code": {
            "type": "string",
            "description": "Python source code to execute."
          },
          "timeout": {
            "type": "number",
            "description": "Maximum execution time in seconds (default: 60).",
            "default": 60
          }
        },
        "required": [
          "code"
        ]
      }
    }
  ],
  "metadata": {
    "task_id": "biostatisticians__task1",
    "reference_files": [
      "dataset/biostatisticians/task1/task_folder/Clinical_Study_Proposal_CVD_Prevention.csv",
      "dataset/biostatisticians/task1/task_folder/framingham.csv"
    ],
    "rubric_json": "{\n  \"rubrics\": [\n    {\n      \"rubric\": \"Does the analysis calculate the observed 10-year CHD event rate in the eligible population as 20.2% (236/1166) and identify this as higher than the assumed 15% control group rate?\",\n      \"weight\": 10,\n ... [TRUNCATED 4642 chars] ... ge criterion and by the sysBP criterion separately\",\n        \"The flow shows the final eligible population count (1,166)\",\n        \"The flow enables verification of the filtering process and assessment of generalizability\"\n      ]\n    }\n  ]\n}"
  }
}

注：部分内容因展示需要已被截断。

提示模板

提示模板：

{question}

使用方法

使用 CLI

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets job_bench \
    --agent-config '{"mode":"native","strategy":"function_calling","max_steps":250}' \
    --limit 10  # 正式评估时请删除此行

使用 Python

from evalscope import TaskConfig, run_task
from evalscope.api.agent import NativeAgentConfig

task_cfg = TaskConfig(
    model='YOUR_MODEL',
    api_url='OPENAI_API_COMPAT_URL',
    api_key='EMPTY_TOKEN',
    datasets=['job_bench'],
    agent_config=NativeAgentConfig(
        strategy='function_calling',
        max_steps=250,
    ),
    limit=10,  # 正式评估时请删除此行
)

run_task(task_cfg=task_cfg)