OpenSportsLab/OSL-VQA-XFOUL-qwen3-8B-VL-lora

OpenSportsLib VQA Model (Qwen3-VL LoRA)

Overview

This model is a VQA LoRA adapter produced with OpenSportsLib for soccer foul understanding and referee-style visual question answering.

• Task: Visual Question Answering (VQA)
• Architecture: Qwen3-VL + LoRA adapter
• Backend: qwen_vl_native_lora
• Base model: Qwen/Qwen3-VL-8B-Instruct
• Library: OpenSportsLib
• Input: Soccer video clips plus natural-language questions
• Visual path: Native end-to-end QwenVL video understanding

Dataset

Training Dataset

This adapter was trained on the OpenSportsLib XFoul VQA setup built on the OSL-XFoul dataset.

• Dataset name: OSL-XFoul
• Domain: Soccer video understanding and officiating
• Task: Visual question answering
• Modality: Video + text
• Training samples: 16,568
• Validation samples: 2,219

Benchmark Results

Accuracy	Balanced Accuracy
69.96%	48.33%

Using with OpenSportsLib

For more details about OpenSportsLib:

• GitHub: https://github.com/OpenSportsLab/opensportslib
• PyPI: https://pypi.org/project/opensportslib/
• Documentation: https://opensportslab.github.io/opensportslib/

Run inference

from opensportslib.apis import VQAModel

my_model = VQAModel(
    config="opensportslib/configs/vqa/qwen3_vl_native.yaml",
    weights="YOUR_HF_REPO_ID",
)

predictions = my_model.infer(
    test_set="/path/to/test_annotations.json",
)

single_prediction = my_model.infer(
    video_path="/path/to/video.mp4",
    question="Was this a foul? What card should be given?",
)

print(predictions)
print(single_prediction)

Notes

• This repository stores a PEFT LoRA adapter, not a merged standalone base model.
• The adapter is intended for the OpenSportsLib native QwenVL VQA path driven by opensportslib/configs/vqa/qwen3_vl_native.yaml.
• The original training setup used frame-based native multimodal inputs with Qwen/Qwen3-VL-8B-Instruct.

License

• Open source license: AGPL 3.0 for research, academic, and community use.
• Commercial license: For proprietary or commercial deployment, please contact the project maintainers.

Citation

@misc{opensportslib_qwen3_vl_vqa_xfoul_lora_2026,
  title={OpenSportsLib Qwen3-VL LoRA for Soccer VQA},
  author={OpenSportsLab},
  year={2026},
  howpublished={https://huggingface.co/OpenSportsLab}
}

Acknowledgements

• Dataset: OpenSportsLab / OSL-XFoul
• Library: https://github.com/OpenSportsLab/opensportslib
• Model pipeline: OpenSportsLib native QwenVL VQA backend