Vision-OPD-Qwen3.5-9B

Vision-OPD Qwen3.5-9B

This is a Vision-OPD fine-tune of Qwen/Qwen3.5-9B, trained with on-policy self-distillation on the Vision-OPD-6K training set.

Training configuration

• Retained training samples after prompt-length filtering: 6,217
• Maximum prompt length: 5,210 tokens
• Maximum response length: 1,024 tokens
• Training epochs: 1
• Optimizer steps: 64
• Learning rate: 2e-6
• Rollouts per prompt: 8
• Teacher regularization: EMA
• Teacher update rate: 0.05

No benchmark evaluation is included with this checkpoint.

Loading

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "YOUR_ORG/Vision-OPD-Qwen3.5-9B"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)

Citation

@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}
}