DefectVideoReward (Qwen3-VL-8B, GRPO step800)

A video defect detection and quality scoring reward model, fine-tuned from Qwen3-VL-8B-Instruct using two-stage training:

1. SFT: Supervised fine-tuning on video defect annotations
2. GRPO: Group Relative Policy Optimization for 800 steps on pairwise video preference data

Task

Given a text prompt and an AI-generated video, the model:

1. Identifies visible defects (artifacts, motion issues, misalignment, etc.)
2. Assigns an overall quality score (1-10)

Input/Output Format

Input: System prompt + user message with video and text prompt

Output: JSON with defects and score

{"defects": [{"issue": "...", "weight": 1|2|3}, ...], "score": 7}

Usage

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor

model = Qwen3VLForConditionalGeneration.from_pretrained(
    "CaoruiLi/DefectVideoReward", torch_dtype="bfloat16", device_map="auto")
processor = AutoProcessor.from_pretrained("CaoruiLi/DefectVideoReward")

Training Details

• Base model: Qwen/Qwen3-VL-8B-Instruct
• SFT data: Video defect annotations with expert-labeled defect lists
• GRPO data: 400 text prompts × pairwise video comparisons from VideoGen-RewardBench
• GRPO steps: 800 (best checkpoint by validation accuracy)
• KL coefficient: 0.01