Awesome

Best Practice

Opencompass VLMEevalKit supports MMT-Bench now! We strongly recommend using VLMEevalKit for its useful features and ready-to-use LVLM implementations.

MMT-Bench

<a href="#🚀-quick-start">Quick Start</a> | <a href="https://mmt-bench.github.io/">HomePage</a> | <a href="https://arxiv.org/abs/2404.16006">arXiv</a> | <a href="https://huggingface.co/datasets/OpenGVLab/MMT-Bench">Dataset</a> | <a href="#🖊️-citation">Citation</a>

This repository is the official implementation of MMT-Bench.

MMT-Bench: A Multimodal MultiTask Benchmark for Comprehensive Evaluation of Large Vision-Language Models
Kaining Ying*, Fanqing Meng*, Jin Wang*, Zhiqian Li, Han Lin, Yue Yang, Hao Zhang, Wenbo Zhang, Yuqi Lin, Shuo Liu, jiayi lei, Quanfeng Lu, Peng Gao, Runjian Chen, Peng Xu, Renrui Zhang, Haozhe Zhang, Yali Wang, Yu Qiao, Ping Luo, Kaipeng Zhang#, Wenqi Shao#
* KY, FM and JW contribute equally.
# WS (shaowenqi@pjlab.org.cn) and KZ (zhangkaipeng@pjlab.org.cn) are correponding authors.

💡 News

2024/07/18: We release the Leaderboard of VAL split. Download dataset here
2024/06/25: We release the ALL split and VAL split.
2024/06/25: The evaluation of ALL split is host on the EvalAI.
2024/06/17: Opencompass VLMEevalKit supports MMT-Bench now! We strongly recommend using VLMEevalKit for its useful features and ready-to-use LVLM implementations.
2024/05/01: MMT-Bench is accepted by ICML 2024. See you in Vienna! 🇦🇹🇦🇹🇦🇹
2024/04/26: We release the evaluation code and the VAL split.
2024/04/24: The technical report of MMT-Bench is released! And check our project page!

Introduction

MMT-Bench is a comprehensive benchmark designed to assess LVLMs across massive multimodal tasks requiring expert knowledge and deliberate visual recognition, localization, reasoning, and planning. MMT-Bench comprises 31, 325 meticulously curated multi-choice visual questions from various multimodal scenarios such as vehicle driving and embodied navigation, covering 32 core meta-tasks and 162 subtasks in multimodal understanding. overview

Evaluation Results Overview

The closed-source proprietary model GPT-4o from OpenAI has taken a leading position in MMT-Bench, surpassing other models such as InternVL-chat, QWen-VL-Plus, GPT-4V, and GeminiProVision. Note that the open-source models InternVL-chat and QwenVL-Max closely follow GPT-4o.
GPT-4o performs well in visual recognition and captioning and improves a lot in visual perception compared with GPT-4V (20231106 & 20240409).

🏆 Leaderboard

Val Set

Rank	Model	Score
1	InternVL2-40B	66.9
2	GPT4o	65.4
3	GeminiPro1-5	64.5
4	GPT4V-20240409-HIGH	64.3
4	InternVL-Chat-V1-2	64.3
6	Claude3-Opus	62.5
7	InternVL2-26B	60.6
8	LLavA-next-Yi-34B	60.4
9	InternVL2-8B	60.0
10	QwenVLMax	59.7
11	GeminiProVision	59.1
12	Mini-InternVL-Chat-4B-V1-5	58.4
13	XComposer2	56.3
14	Yi-VL-6B	54.7
15	Phi-3-Vision	54.5
15	TransCore-M	54.5
17	deepseek-vl-7B	54.0
17	Yi-VL-34B	54.0
19	LLavA-internlm2-7B	53.4
19	Monkey-Chat	53.4
21	LLavA-next-vicuna-13B	52.4
22	LLavA-v1.5-13B	52.1
23	sharegpt4v-7B	51.6
24	LLavA-v1.5-13B-xtuner	50.7
25	mPLUG-Owl2	50.5
26	LLavA-next-vicuna-7B	50.4
27	LLavA-v1.5-7B	49.6
28	LLavA-v1.5-7B-xtuner	49.3
29	LLavA-internlm-7B	48.3
30	Qwen-Chat	47.9
30	sharecaptioner	47.9

Full Set

Rank	Model	Score
1	GPT4o	65.5
2	InternVL-Chat-v1.2-34B	63.4
3	QwenVLMax	62.4
4	Qwen-VL-Plus	62.3
5	GeminiProVision	61.6
6	GPT4V_20240409	61.1
7	LLaVA-NEXT-34B	60.8
8	XComposer2	55.7
9	BLIP2	54.8
10	GPT4V_20231106	54.7
11	Yi-VL-34B	54.2
12	Monkey-Chat	53.4
13	DeepSeek-VL-7B	53.2
14	Yi-VL-6B	53.2
15	LLaVA-NEXT-13B	53.0
16	TransCore-M	52.7
17	QWen-VL-Chat	52.5
18	Claude3V_Haiku	52.2
19	XComposer	52.1
20	mPLUG-Owl2	52.0
21	RBDash-v1-13B	51.8
22	LLaVA-v1.5-13B	51.7
23	CogVLM-Chat	51.6
24	ShareGPT4V-7B	51.5
25	LLaVA-NEXT-7B	51.1
26	LLaVA-v1.5-13B-XTuner	51.1
27	LLaVA-InternLM2-7B	50.8
28	LLaVA-v1.5-7B-XTuner	50.2
29	SharedCaptioner	49.9
30	LLaVA-InternLM-7B	49.7
31	LLaVA-v1.5-7B	49.5
32	LLaMA-Adapter-v2-7B	40.4
33	VisualGLM-6B	38.6
34	Frequency Guess	31.7
35	Random Guess	28.5

🚀 Quick Start

Please refer to this to quick start.

💐 Acknowledgement

We expressed sincerely gratitude for the projects listed following:

VLMEvalKit provides useful out-of-box tools and implements many adavanced LVLMs. Thanks for their selfless dedication.

🖊️ Citation

If you feel MMT-Bench useful in your project or research, please kindly use the following BibTeX entry to cite our paper. Thanks!

@misc{mmtbench,
    title={MMT-Bench: A Comprehensive Multimodal Benchmark for Evaluating Large Vision-Language Models Towards Multitask AGI}, 
    author={Kaining Ying and Fanqing Meng and Jin Wang and Zhiqian Li and Han Lin and Yue Yang and Hao Zhang and Wenbo Zhang and Yuqi Lin and Shuo Liu and Jiayi Lei and Quanfeng Lu and Runjian Chen and Peng Xu and Renrui Zhang and Haozhe Zhang and Peng Gao and Yali Wang and Yu Qiao and Ping Luo and Kaipeng Zhang and Wenqi Shao},
    year={2024},
    eprint={2404.16006},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}