dataset9/README.md

---
license: apache-2.0
base_model:
- Qwen/Qwen3.5-9B
language:
- en
- zh
library_name: transformers
pipeline_tag: image-text-to-text
datasets:
- yuanqianhao/Vision-OPD-6K
tags:
- multimodal
- vision-language
- fine-grained-visual-understanding
- on-policy-distillation
- self-distillation
---

# Vision-OPD-9B

<p align="center">
  📄 <a href="https://arxiv.org/abs/2605.18740">Paper</a> &nbsp;|&nbsp;
  💻 <a href="https://github.com/VisionOPD/Vision-OPD">Code</a> &nbsp;|&nbsp;
  🤗 <a href="https://huggingface.co/datasets/yuanqianhao/Vision-OPD-6K">Training Data</a> &nbsp;|&nbsp;
  📦 <a href="https://hf.co/collections/yuanqianhao/vision-opd">Collection</a>
</p>

## Model Summary

**Vision-OPD-9B** is a multimodal LLM built on [`Qwen/Qwen3.5-9B`](https://huggingface.co/Qwen/Qwen3.5-9B) and trained with
**Vision-OPD (Vision On-Policy Distillation)**, the regional-to-global self-distillation framework introduced in
[Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation](https://arxiv.org/abs/2605.18740).

MLLMs often answer fine-grained questions correctly when given an evidence-centered **crop**, yet fail on the corresponding
**full image** — a persistent *regional-to-global perception gap*. Vision-OPD closes this gap by letting the model act as its
own teacher: a **crop-conditioned teacher** and a **full-image-conditioned student** are instantiated from the same MLLM, the
student generates on-policy rollouts, and training minimizes the token-level divergence between the two distributions along
these rollouts. The benefit of visual zooming is thereby internalized into a **single forward pass** — no external teacher,
no ground-truth labels, no reward verifiers, and no inference-time tool use.

With only **6.2K synthetic samples**, Vision-OPD-9B outperforms much larger open-source models (e.g., Qwen3.5-397B, Kimi-K2.6),
closed-source models (e.g., GPT-5.4, Gemini-3.1-Pro), and "Thinking-with-Images" agentic methods on fine-grained visual
understanding benchmarks, while preserving general multimodal abilities on holdout tasks.

## Key Features

- 🔍 **Sees fine details in one forward pass** — no cropping tools, zoom-in calls, or multi-step agentic inference at test time.
- 🪞 **Self-distillation, label-free** — the model's own crop-conditioned perception supervises its full-image policy; no external teacher, ground-truth labels, or reward verifiers.
- 🎯 **On-policy + dense supervision** — token-level divergence (JSD, β = 0.5) on the student's own rollouts avoids the exposure bias of SFT and the sparse rewards of RLVR.
- 📉 **Data-efficient** — trained on only 6.2K automatically synthesized samples ([Vision-OPD-6K](https://huggingface.co/datasets/yuanqianhao/Vision-OPD-6K)), one epoch.
- 🧠 **No forgetting** — general multimodal capability is preserved on holdout tasks (MMVP, CV-Bench, MMStar, POPE).

## How It Works

1. **Data synthesis** — evidence regions are proposed on unlabeled images via object identification and segmentation; a question answerable from the crop alone is generated, and the region's bounding box is overlaid on the full image for grounding. This yields triplets *(full image, crop, question)*.
2. **Two policies, one model** — the teacher conditions on the privileged crop, the student on the full image.
3. **On-policy distillation** — the student samples a rollout (up to 1024 tokens); at every position both policies are evaluated on the same student prefix, and the per-token JSD between them is minimized (top-K = 100 logits distillation). Gradients flow only through the student.
4. **EMA teacher regularization** — the teacher is updated as an exponential moving average of the student (α = 0.05), which prevents teacher–student co-adaptation and training collapse.

## Quickstart

Weights load with standard `transformers` (≥ 5.5):

```python
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "yuanqianhao/Vision-OPD-9B"
model = AutoModelForImageTextToText.from_pretrained(model_id, dtype="bfloat16", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id)

messages = [{
    "role": "user",
    "content": [
        {"type": "image", "url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"},
        {"type": "text", "text": "What is the color of the label on the bottle in the background?"},
    ],
}]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(processor.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
```

Or serve with vLLM (OpenAI-compatible API), as used in our evaluation:

```bash
vllm serve yuanqianhao/Vision-OPD-9B --port 8000
```

To reproduce the reported numbers, use the released evaluation harness:

```bash
git clone https://github.com/VisionOPD/Vision-OPD && cd Vision-OPD
bash eval/run_eval.sh   # see README for API base / judge / benchmark env vars
```

## Training Details

| | |
|---|---|
| Base model | [`Qwen/Qwen3.5-9B`](https://huggingface.co/Qwen/Qwen3.5-9B) (non-thinking mode) |
| Data | [Vision-OPD-6K](https://huggingface.co/datasets/yuanqianhao/Vision-OPD-6K) — 6.2K synthesized (full image, crop, question) triplets |
| Objective | Token-level JSD (β = 0.5) on on-policy student rollouts, top-K = 100 logits distillation |
| Teacher | Same model conditioned on the privileged crop, EMA-regularized (α = 0.05) |
| Rollout length | 1024 tokens, 1 epoch |

Training code: [github.com/VisionOPD/Vision-OPD](https://github.com/VisionOPD/Vision-OPD).

## Citation

```bibtex
@article{yuan2026vision,
  title={Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation},
  author={Yuan, Qianhao and Lou, Jie and Yu, Xing and Lin, Hongyu and Sun, Le and Han, Xianpei and Lu, Yaojie},
  journal={arXiv preprint arXiv:2605.18740},
  year={2026}
}
```

## License

Released under Apache-2.0, following the base model [`Qwen/Qwen3.5-9B`](https://huggingface.co/Qwen/Qwen3.5-9B).