array/Qwen2.5-VL-SAT

Model Card for Model ID

A strong spatial Qwen 2.5 VL baseline.

Post-trained on SAT, and just the answers from Video-R1. The exact mix is 60% SAT, 40% Video-R1.

Get Started

% pip install git+https://github.com/huggingface/transformers accelerate
% pip install qwen-vl-utils[decord]==0.0.8

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor

model = Qwen2_5_VLForConditionalGeneration.from_pretrained("array/Qwen2.5-VL-SAT")
processor = AutoProcessor.from_pretrained(
    exp_confs["model_path"],
    trust_remote_code=model_config.trust_remote_code
)

Please see the paper for details on training and evaluation datasets and metrics.

Results

Model	MV	RelDep	SpRel	Jig	IQT	BLINK Avg	BLINK Reas	SAT-R	VSI Avg	VSI Reas	ERQA	Avg (All)
Qwen2.5-VL (7B)	39.00	61.29	92.38	58.66	25.33	55.33	41.00	59.00	23.96	22.96	38.91	44.30
+ SAT	57.14	87.09	74.12	58.66	30.00	61.40	48.60	71.66	32.40	30.65	38.00	50.87

Citation [optional]

@misc{ray2025satdynamicspatialaptitude,
      title={SAT: Dynamic Spatial Aptitude Training for Multimodal Language Models}, 
      author={Arijit Ray and Jiafei Duan and Ellis Brown and Reuben Tan and Dina Bashkirova and Rose Hendrix and Kiana Ehsani and Aniruddha Kembhavi and Bryan A. Plummer and Ranjay Krishna and Kuo-Hao Zeng and Kate Saenko},
      year={2025},
      eprint={2412.07755},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2412.07755}, 
    }