Model Details

• Model type: no-reference video quality assessment vision-language model
• Checkpoint type: PEFT / LoRA adapter for active fine-tuning
• Backbone family: Qwen2.5-VL / VisualQuality-R1-style VLM
• Base model: hollow404/VQR1-7B-YouTubeUGC
• LoRA rank: 64
• LoRA alpha: 128
• LoRA dropout: 0.05
• Training data: YouTube-UGC + MDS-VQA-selected labeled samples from YouTube-SFV SDR
• Input: a video plus a VQA prompt
• Output: a quality score on a 1 to 5 scale, typically inside <answer>...</answer> tags
• License: Apache 2.0

Intended Use

This model is intended for research on no-reference video quality assessment, active data selection, and target-domain adaptation for VQA. Typical uses include:

• evaluating the active fine-tuning stage of the MDS-VQA pipeline;
• predicting perceptual quality scores for YouTube-SFV SDR;
• comparing active fine-tuning against the YouTube-UGC baseline model;
• studying how model-informed data selection improves VQA generalization.

This checkpoint should be used together with the base model. It is not intended as a universal production QoE monitor without domain-specific validation.

Prompt Format

The model follows the VisualQuality-R1-style scoring prompt used in MDS-VQA:

text
Copy code
You are doing the video quality assessment task.
Here is the question: What is your overall rating on the quality of this video? The rating should be a float between 1 and 5, rounded to two decimal places, with 1 representing very poor quality and 5 representing excellent quality.
First output the thinking process in <think> </think> tags and then output the final answer with only one score in <answer> </answer> tags.

For automatic evaluation, parse the scalar value inside the final <answer> tag.

MDS-VQA Context

MDS-VQA is a model-informed data selection mechanism for VQA. Given an unlabeled target video pool, it selects videos that are both:

Difficult for the base VQA model: estimated by a failure predictor trained to rank videos by the base model's prediction errors. Diverse in content: estimated from semantic video features, using a diversity-aware greedy selection procedure. The selected videos are then labeled and merged with the original labeled source dataset for active fine-tuning. This repository provides the resulting active fine-tuning checkpoint.

Citation

If you use this model, please cite MDS-VQA:

bibtext
Copy code
@article{zou2026mds,
  title={MDS-VQA: Model-Informed Data Selection for Video Quality Assessment},
  author={Zou, Jian and Xu, Xiaoyu and Wang, Zhihua and Wang, Yilin and Adsumilli, Balu and Ma, Kede},
  journal={arXiv preprint arXiv:2603.11525},
  year={2026}
}