Vision-OPD-6K: Training Data for Vision-OPD

Training data for Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation.

Overview

Vision-OPD proposes a regional-to-global self-distillation framework that transfers the model's own privileged regional perception to its full-image policy, without external teacher models, ground-truth labels, reward verifiers, or inference-time tool use.

Vision-OPD instantiates two conditional policies from the same MLLM:

• A crop-conditioned teacher that observes the evidence-centered crop as a privileged input
• A full-image-conditioned student that observes the full image as in standard inference

The student generates on-policy rollouts, and Vision-OPD minimizes token-level divergence between the teacher and student next-token distributions along these rollouts.

This dataset contains 6.2K fine-grained visual question answering samples used for Vision-OPD training. Each sample includes a full image with red bounding box overlay (student input) and a cropped region image (teacher input).

Dataset Statistics

Attribute	Value
Total Samples	6,241
Image Types	Full image with bbox overlay (student) + cropped region (teacher)

Data Structure & Extraction

The image data is provided in compressed files.

1. Extract Student Images

cd images/
cat images.tar.gz* | tar -xvf - -C ./

2. Extract Teacher Images

cd teacher_images/
tar -xvf teacher_images.tar.gz -C ./

Data Format

Each record in train.jsonl contains:

Field	Type	Description
images	list[str]	Relative path to bbox overlay image (student input)
teacher_images	list[str]	Relative path to cropped region image (teacher input)
problem	str	Question prompt with <image> token
answer	str	Ground truth answer
extra_info	dict	Contains answer and question (without <image> prefix)

Training Results

Models trained with Vision-OPD on this dataset achieve competitive or superior performance against much larger open-source, closed-source, and "Thinking-with-Images" agentic models.

Comparison with SOTA MLLMs

Model	Param	V* Bench	ZoomBench	HR Bench 4K	HR Bench 8K	MME-RW Lite	MME-RW CN	Average
DeepEyes	7B	85.86	46.51	75.13	72.63	53.20	64.09	66.24
GPT-5.2	-	79.06	50.89	81.12	78.38	47.06	68.80	67.55
GPT-5.4	-	76.96	52.66	84.00	77.88	48.10	70.93	68.42
Qwen3.5	397B	87.96	57.16	89.38	85.50	45.39	69.82	72.54
Gemini-3.1-Pro	-	87.96	61.18	89.63	86.88	49.50	73.31	74.74
Vision-OPD (Ours)	4B	92.15	59.76	84.50	80.38	55.45	70.76	73.83
Vision-OPD (Ours)	9B	94.76	65.80	88.13	85.50	49.56	70.46	75.70

Intended Use

This dataset is designed for:

• On-policy self-distillation training of MLLMs for fine-grained visual understanding
• Research on regional-to-global perception transfer without inference-time tool use

Related Resources

Resource	Link
Paper	arXiv:2605.18740
Code	GitHub: VisionOPD/Vision-OPD

Citation

@misc{yuan2026visionopdlearningfinedetails,
      title={Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation}, 
      author={Qianhao Yuan and Jie Lou and Xing Yu and Hongyu Lin and Le Sun and Xianpei Han and Yaojie Lu},
      year={2026},
      eprint={2605.18740},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.18740}, 
}

License

Apache-2.0 License