evalstone/evalscope/docs/en/benchmarks/one_million_bench.md

$OneMillion-Bench

Overview

$OneMillion-Bench ($1M-Bench) evaluates how well language models and agents complete economically valuable, expert-level professional work. The public release contains 400 bilingual tasks written and reviewed by domain experts across finance, healthcare, industry, law, and natural science.

Task Description

• Task Type: Open-ended professional question answering with rubric-based LLM judging
• Input: A realistic, context-heavy professional request in Chinese or English
• Output: A complete free-form professional analysis or deliverable
• Domain: Economics and finance, healthcare and medicine, industry, law, and natural sciences

Key Features

• 400 zero-shot tasks, balanced across Chinese and global tracks and five professional domains (40 tasks per language-domain subset)
• Each task has 11-37 expert-authored criteria covering factual information, analytical reasoning, instruction following, and structure and formatting
• Every task includes both positive criteria and negative penalties, with rubric weights ranging from -20 to 12 in the hosted release
• Samples are exposed as ten language-domain subsets so both the paper's language tracks and domain breakdowns are visible in EvalScope reports

Evaluation Notes

• An LLM judge is required. Configure judge.strategy='llm' (or 'auto') and judge.models; the official harness currently recommends Gemini 3.1 Pro Preview, but judge identity affects absolute scores and is not hard-coded here
• All rubrics for one response are judged together in one request using the official binary hit/miss instructions
• expert_score is the weighted sum of hit rubrics divided by the sum of positive weights, clipped to [0, 1]; pass_rate is 1 when expert_score >= 0.7, otherwise 0
• Judge replies must contain every rubric exactly once. Malformed replies and transport failures are excluded instead of being silently converted to zero scores
• The official study compares vanilla models, search-enabled models, and deep-research agents separately. This native adapter performs the benchmark's one-turn generation path; results from external tool-using agents are comparable only when their final responses are evaluated under the same judge configuration
• Tasks often require long, cited reports. Configure sufficiently large generation and judge max_tokens values; with one judge and one repeat, a full run performs 400 generation calls and 400 judge calls

Resources: Paper | GitHub | Dataset

Properties

Property	Value
Benchmark Name	one_million_bench
Dataset ID	evalscope/OneMillion-Bench
Paper	Paper
Tags	Agent, Knowledge, MultiLingual, QA, Reasoning
Metrics	expert_score, pass_rate
Default Shots	0-shot
Evaluation Split	test

Data Statistics

Metric	Value
Total Samples	400
Prompt Length (Mean)	1470.49 chars
Prompt Length (Min/Max)	105 / 15951 chars

Per-Subset Statistics:

Subset	Samples	Prompt Mean	Prompt Min	Prompt Max
global_economics_and_finance	40	1810.97	868	3723
global_healthcare_and_medicine	40	2029.97	479	8870
global_industry	40	2581.15	453	9538
global_law	40	3277.78	404	15951
global_natural_sciences	40	1634.65	284	5775
cn_economics_and_finance	40	555.05	111	1169
cn_healthcare_and_medicine	40	584.38	135	2628
cn_industry	40	965.6	144	7755
cn_law	40	709.55	208	2041
cn_natural_sciences	40	555.83	105	1590

Sample Example

Subset: global_economics_and_finance

{
  "input": [
    {
      "id": "24273bb5",
      "content": "You are an international financial risk analyst. Based on the Financial Stability Report released by the Bank of England's Financial Policy Committee in October 2025, global financial markets may face a \"risk of sharp market correction\" if in ... [TRUNCATED 924 chars] ... ields, and global capital flows.\nSpecial Conditions: Use only information published before December 31, 2025; do not fabricate information; generated content must cite real URLs. The answer must be complete and useful; do not fake a response."
    }
  ],
  "target": "[{\"rubric_number\": 1, \"rubric_detail\": \"Mention the high weighting of top US companies (e.g., Top 5 or AI-related stocks) in the index (approximately 30% or more) and identify this as a source of systemic risk.\", \"rubric_weight\": 10, \"rubric_ ... [TRUNCATED 3986 chars] ... c_number\": 16, \"rubric_detail\": \"The report contains hollow concluding remarks (e.g., \\\"In summary,\\\" \\\"We look forward to\\\") or transitional sentences lacking substantive content.\", \"rubric_weight\": -2, \"rubric_tag\": \"Analytical Reasoning\"}]",
  "id": 0,
  "group_id": 0,
  "subset_key": "global_economics_and_finance",
  "metadata": {
    "id": "e1b94c86-b6c9-43f6-8251-2e513e5efc52",
    "case_id": 1663,
    "language": "global",
    "domain": "Economics and Finance",
    "topics": [
      "Economics and Finance",
      "Financing & M&A",
      "Mergers & Acquisitions"
    ],
    "time_sensitivity": {
      "time_sensitivity": "Weakly time-sensitive",
      "year_month": "2025-10",
      "day": "NA"
    },
    "question": "You are an international financial risk analyst. Based on the Financial Stability Report released by the Bank of England's Financial Policy Committee in October 2025, global financial markets may face a \"risk of sharp market correction\" if in ... [TRUNCATED 924 chars] ... ields, and global capital flows.\nSpecial Conditions: Use only information published before December 31, 2025; do not fabricate information; generated content must cite real URLs. The answer must be complete and useful; do not fake a response."
  }
}

Note: Some content was truncated for display.

Prompt Template

Prompt Template:

{question}

Usage

Using CLI

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets one_million_bench \
    --limit 10  # Remove this line for formal evaluation

Using Python

from evalscope import run_task
from evalscope.config import TaskConfig

task_cfg = TaskConfig(
    model='YOUR_MODEL',
    api_url='OPENAI_API_COMPAT_URL',
    api_key='EMPTY_TOKEN',
    datasets=['one_million_bench'],
    dataset_args={
        'one_million_bench': {
            # subset_list: ['global_economics_and_finance', 'global_healthcare_and_medicine', 'global_industry']  # optional, evaluate specific subsets
        }
    },
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)