BAAI/Aquila-VL-2B-llava-qwen

Introduction

The Aquila-VL-2B model is a vision-language model (VLM) trained based on the LLava-one-vision framework. The Qwen2.5-1.5B-instruct model is chose as the LLM, while siglip-so400m-patch14-384 is utilized as the vision tower.

The model was trained on our self-built Infinity-MM dataset, which contains approximately 40 million image-text pairs. This dataset is a combination of open-source data collected from the internet and synthetic instruction data generated using open-source VLM models.

We have open-sourced Infinity-MM dataset and related resources. We hope you enjoy using them!

News

• 2024/10/25: The Aquila-VL-2B model and Infinity-MM dataset are now available. We have also released the technical report simultaneously.

Evaluation

We evaluated the model using the VLMEvalKit tool. Whenever possible, we prioritized using the OpenAI API for test sets that support API-based evaluation.

Benchmark	MiniCPM-V-2	InternVL2-2B	XinYuan-VL-2B	Qwen2-VL-2B-Instruct	Aquila-VL-2B
MMBench-ENtest	69.4	73.4	78.9	74.9	78.8
MMBench-CNtest	65.9	70.9	76.1	73.9	76.4
MMBench_V1.1test	65.2	69.7	75.4	72.7	75.2
MMT-Benchtest	54.5	53.3	57.2	54.8	58.2
RealWorldQA	55.4	57.3	63.9	62.6	63.9
HallusionBench	36.8	38.1	36.0	41.5	43.0
SEEDBench2plus	51.8	60.0	63.0	62.4	63.0
LLaVABench	66.1	64.8	42.4	52.5	68.4
MMStar	41.6	50.2	51.9	47.8	54.9
POPE	86.6	85.3	89.4	88.0	83.6
MMVet	44.0	41.1	42.7	50.7	44.3
MMMUval	39.6	34.9	43.6	41.7	47.4
ScienceQAtest	80.4	94.1	86.6	78.1	95.2
AI2Dtest	64.8	74.4	74.2	74.6	75.0
MathVistatestmini	39.0	45.0	47.1	47.9	59.0
MathVersetestmini	19.8	24.7	22.2	21.0	26.2
MathVision	15.4	12.6	16.3	17.5	18.4
DocVQAtest	71.0	86.9	87.6	89.9	85.0
InfoVQAtest	40.0	59.5	59.1	65.4	58.3
ChartQAtest	59.6	71.4	57.1	73.5	76.5
TextVQAval	74.3	73.5	77.6	79.9	76.4
OCRVQAtestcore	54.4	40.2	67.6	68.7	64.0
VCRen easy	27.6	51.6	67.7	68.3	70.0
OCRBench	613	784	782	810	772
Average	53.5	58.8	60.9	62.1	64.1

For comparison models, evaluations were conducted in a local environment, so the scores may differ slightly from those reported in papers or on the official VLMEvalKit leaderboard.

Future Plan

• We plan to train models of various sizes.
• Future training will incorporate multi-image and video data.

Citation

If you find this dataset useful, please cite the following work

@misc{gu2024infinitymmscalingmultimodalperformance,
      title={Infinity-MM: Scaling Multimodal Performance with Large-Scale and High-Quality Instruction Data}, 
      author={Shuhao Gu and Jialing Zhang and Siyuan Zhou and Kevin Yu and Zhaohu Xing and Liangdong Wang and Zhou Cao and Jintao Jia and Zhuoyi Zhang and Yixuan Wang and Zhenchong Hu and Bo-Wen Zhang and Jijie Li and Dong Liang and Yingli Zhao and Yulong Ao and Yaoqi Liu and Fangxiang Feng and Guang Liu},
      year={2024},
      eprint={2410.18558},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.18558}, 
}