MMBench#

Overview#

MMBench is a systematically designed benchmark for evaluating vision-language models across 20 fine-grained ability dimensions. It uses a novel CircularEval strategy and provides both English and Chinese versions for cross-lingual evaluation.

Task Description#

Task Type: Visual Multiple-Choice Q&A
Input: Image with question and 2-4 answer options
Output: Single correct answer letter (A, B, C, or D)
Languages: English (en) and Chinese (cn) subsets

Key Features#

20 fine-grained ability dimensions defined
Systematic evaluation pipeline with CircularEval strategy
Bilingual support (English and Chinese)
Questions include hints for context
Tests perception, reasoning, and knowledge abilities

Evaluation Notes#

Default configuration uses 0-shot evaluation
Uses Chain-of-Thought (CoT) prompting
Evaluates on dev split
Two subsets: cn (Chinese) and en (English)
Results include category-level breakdown

Properties#

Property	Value
Benchmark Name	`mm_bench`
Dataset ID	lmms-lab/MMBench
Paper	N/A
Tags	`Knowledge`, `MultiModal`, `QA`
Metrics	`acc`
Default Shots	0-shot
Evaluation Split	`dev`

Data Statistics#

Metric	Value
Total Samples	8,658
Prompt Length (Mean)	372.3 chars
Prompt Length (Min/Max)	241 / 2395 chars

Per-Subset Statistics:

Subset	Samples	Prompt Mean	Prompt Min	Prompt Max
`cn`	4,329	312.28	241	1681
`en`	4,329	432.33	256	2395

Image Statistics:

Metric	Value
Total Images	8,658
Images per Sample	min: 1, max: 1, mean: 1
Resolution Range	106x56 - 512x512
Formats	jpeg

Sample Example#

Subset: cn

{
  "input": [
    {
      "id": "00b49734",
      "content": [
        {
          "text": "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. Think step by step before answering.\n\n下面的文章描述了一个实验。阅读文章，然后按照以下说明进行操作。\n\nMadelyn在雪板的底部涂上了一层薄蜡，然后直接下坡滑行。然后，她去掉了蜡，再次直接下坡滑行。她重复了这个过程四次，每次都交替使用薄蜡或不使用薄蜡滑行。她的朋友Tucker计时每次滑行的时间。Madelyn和Tucker计算了使用薄蜡滑行和不使用薄蜡滑行时直接下坡所需的平均时间。\n图：滑雪板下坡。麦德琳和塔克的实验能最好回答哪个问题？\n\nA) 当麦德琳的雪板上有一层薄蜡或一层厚蜡时，它是否能在较短的时间内滑下山坡？\nB) 当麦德琳的雪板上有一层蜡或没有蜡时，它是否能在较短的时间内滑下山坡？"
        },
        {
          "image": "[BASE64_IMAGE: jpeg, ~11.7KB]"
        }
      ]
    }
  ],
  "choices": [
    "当麦德琳的雪板上有一层薄蜡或一层厚蜡时，它是否能在较短的时间内滑下山坡？",
    "当麦德琳的雪板上有一层蜡或没有蜡时，它是否能在较短的时间内滑下山坡？"
  ],
  "target": "B",
  "id": 0,
  "group_id": 0,
  "metadata": {
    "index": 241,
    "category": "identity_reasoning",
    "source": "scienceqa",
    "L2-category": "attribute_reasoning",
    "comment": "nan",
    "split": "dev"
  }
}

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 {letters}. Think step by step before answering.

{question}

{choices}

Usage#

Using CLI#

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets mm_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=['mm_bench'],
    dataset_args={
        'mm_bench': {
            # subset_list: ['cn', 'en']  # optional, evaluate specific subsets
        }
    },
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)