MMStar

Overview

MMStar is an elite vision-indispensable multimodal benchmark designed to ensure genuine visual dependency in evaluation. Each sample is carefully curated to require actual visual understanding, minimizing data leakage and testing advanced multimodal capabilities.

Task Description

• Task Type: Vision-Dependent Multiple-Choice QA

• Input: Image + multiple-choice question requiring visual understanding

• Output: Single answer letter (A/B/C/D)

• Domains: Perception, reasoning, math, science & technology

Key Features

• Ensures visual dependency - questions cannot be answered without images

• Minimal data leakage from training corpora

• Tests advanced multimodal reasoning capabilities

• Six categories: coarse perception, fine-grained perception, instance reasoning, logical reasoning, math, science & technology

• High-quality curated samples with verified visual necessity

Evaluation Notes

• Default evaluation uses the val split

• Primary metric: Accuracy on multiple-choice questions

• Uses Chain-of-Thought (CoT) prompting with “ANSWER: [LETTER]” format

• Results reported per category and overall

Properties

Property	Value
Benchmark Name	mm_star
Dataset ID	evalscope/MMStar
Paper	N/A
Tags	Knowledge, MCQ, MultiModal
Metrics	acc
Default Shots	0-shot
Evaluation Split	val

Data Statistics

Metric	Value
Total Samples	1,500
Prompt Length (Mean)	390.23 chars
Prompt Length (Min/Max)	272 / 2023 chars

Per-Subset Statistics:

Subset	Samples	Prompt Mean	Prompt Min	Prompt Max
coarse perception	250	350.8	282	784
fine-grained perception	250	334.98	277	608
instance reasoning	250	379.02	273	684
logical reasoning	250	427.22	284	2023
math	250	467.36	292	891
science & technology	250	381.98	272	1173

Image Statistics:

Metric	Value
Total Images	1,500
Images per Sample	min: 1, max: 1, mean: 1
Resolution Range	114x66 - 3160x2136
Formats	jpeg

Sample Example

Subset: coarse perception

{
  "input": [
    {
      "id": "57e256b8",
      "content": [
        {
          "text": "Answer the following multiple choice question.\nThe last line of your response should be of the following format:\n'ANSWER: [LETTER]' (without quotes)\nwhere [LETTER] is one of A,B,C,D. Think step by step before answering.\n\nWhich option describe the object relationship in the image correctly?\nOptions: A: The suitcase is on the book., B: The suitcase is beneath the cat., C: The suitcase is beneath the bed., D: The suitcase is beneath the book."
        },
        {
          "image": "[BASE64_IMAGE: jpeg, ~37.2KB]"
        }
      ]
    }
  ],
  "choices": [
    "A",
    "B",
    "C",
    "D"
  ],
  "target": "A",
  "id": 0,
  "group_id": 0,
  "subset_key": "coarse perception",
  "metadata": {
    "index": 0,
    "category": "coarse perception",
    "l2_category": "image scene and topic",
    "source": "MMBench",
    "split": "val",
    "image_path": "images/0.jpg"
  }
}

Prompt Template

Prompt Template:

Answer the following multiple choice question.
The last line of your response should be of the following format:
'ANSWER: [LETTER]' (without quotes)
where [LETTER] is one of A,B,C,D. Think step by step before answering.

{question}

Usage

Using CLI

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets mm_star \
    --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=['mm_star'],
    dataset_args={
        'mm_star': {
            # subset_list: ['coarse perception', 'fine-grained perception', 'instance reasoning']  # optional, evaluate specific subsets
        }
    },
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)