MedM-VL-2D-3B-en

Introduction

A 2D medical LVLM trained on 2D medical images and English medical texts, enabling tasks such as report generation, VQA, referring expression comprehension (REC), referring expression generation (REG) and image classification.

	Config
Image encoder	google/siglip-base-patch16-256-multilingual
Connector	MLP (2-layer)
LLM	Qwen/Qwen2.5-3B-Instruct
Image resolution	256*256
Sequence length	2048

Evaluation

Benchmark	Med-Flamingo	LLaVA-Med	RadFM	MedM-VL-2D-3B-en
MedMNISTderma	0.012	0.258	0.051	0.786
MedMNISTorgan	0.089	0.668	0.189	0.808
MedPix	0.081	0.151	-	0.126
MIMIC-CXR	0.233	0.204	0.068	0.199
PathVQA	0.334	0.378	0.248	0.634
SAMedidentify	-	0.458	-	0.693
SAMedrefer	-	0.086	-	0.235
SLAKEidentify	-	0.272	-	0.727
SLAKErefer	-	0.041	-	0.313
SLAKEvqa	0.215	0.337	0.817	0.841

Quickstart

Please refer to MedM-VL.

Citation

@article{shi2025medm,
  title={MedM-VL: What Makes a Good Medical LVLM?},
  author={Shi, Yiming and Yang, Shaoshuai and Zhu, Xun and Wang, Haoyu and Li, Miao and Wu, Ji},
  journal={arXiv preprint arXiv:2504.04323},
  year={2025}
}