Modelpublic

leoyinn/qwen3vl-flare25

sourceHugging Faceapache-2.0updated 11mo agoView on Hugging Face

1likes136downloads

Model Card

Qwen3-VL-4B-FLARE25

Model Description

This is a fully finetuned version of Qwen/Qwen3-VL-4B-Instruct on the FLARE 2025 medical imaging dataset. The model has been trained to perform various medical vision-language tasks across 8 imaging modalities and 19 different datasets.

Base Model: Qwen3-VL-4B-Instruct Training Dataset: FLARE25 (Medical Imaging Foundation Models: A Multi-task Learning Framework) Training Samples: 12232 samples across 26 datasets Supported Tasks:

—Classification
—Multi-label Classification
—Detection
—Instance Detection
—Regression
—Counting
—Report Generation

Supported Medical Imaging Modalities

1.Ultrasound - Breast ultrasound, intrauterine growth charts
2.X-ray - Dental, chest, periapical radiographs
3.Retinography - Fundus imaging, diabetic retinopathy
4.Microscopy - Chromosome analysis, bone marrow, cell counting
5.Clinical Photography - Neonatal jaundice assessment
6.Dermatology - Skin lesion classification
7.Endoscopy - Gastrointestinal imaging
8.Mammography - Breast cancer screening

Performance Summary

Task	Primary Metric	Baseline	Finetuned	Improvement
Classification	Balanced Accuracy	2.2%	53.5%	+2,309%
Detection	F1@0.5	0.0%	80.3%	∞ (new capability)
Instance Detection	F1@0.5	0.01%	1.0%	+9,900%
Multi-label Classification	F1 Macro	28.3%	50.3%	+77.7%
Regression	MAE	35.8	22.4	+37.3%
Counting	MAE	417.7	244.4	+41.5%
Report Generation	GREEN Score	67.7%	80.8%	+19.4%

Usage

Please check our GitHub for details

Training Details

Training Hyperparameters

—Base Model: Qwen3-VL-4B-Instruct
—Training Framework: DeepSpeed ZeRO-3
—Learning Rate: 1e-5
—Batch Size: 4 per device
—Gradient Accumulation Steps: 4
—Training Epochs: 2
—Max Sequence Length: 8192
—Image Resolution: Dynamic (maxpixels: 50176, minpixels: 784)
—Optimizer: AdamW
—Mixed Precision: BF16
—Gradient Checkpointing: Enabled

Training Data Distribution

The model was trained on 19 medical imaging datasets across 8 modalities:

Ultrasound:

—BUSI, BUS-UCLM (Classification)
—BUSI-det, BUS-UCLM-det (Detection)
—IUGC (Classification + Detection)

X-ray:

—Dental, Periapical, Bone Resorption, ChestDR, IU-XRay

Retinography:

—Retino, Fundus

Microscopy:

—Chromosome, Bone Marrow, NeurIPS22-Cell

Clinical/Dermatology/Endoscopy/Mammography:

—Neojaundice, BCN20000, Endo, CMMD

Citation

If you use this model, please cite:

bibtex

@misc{qwen3vl-flare25,
  author = {Shuolin Yin},
  title = {Qwen3-VL-4B Finetuned on FLARE25 Medical Imaging Dataset},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/leoyinn/qwen3vl-flare25}}
}

@misc{qwen3technicalreport,
      title={Qwen3 Technical Report}, 
      author={Qwen Team},
      year={2025},
      eprint={2505.09388},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.09388}, 
}

License

This model is released under the Apache 2.0 License. The base model license from Qwen also applies.

Acknowledgments

—Base Model: Qwen3-VL-4B-Instruct by Alibaba Cloud
—Dataset: FLARE 2025 Medical Imaging Challenge
—Training Infrastructure: Built on the official Qwen3-VL finetuning framework

Repository

Full training code, evaluation scripts, and results: GitHub - FLARE25-QWen3VL-4B