Qwen2.5-7B-VL-ReAd-R
ReAd-R is a Qwen2.5-VL-7B based video understanding model fine-tuned with GRPO reinforcement learning for advertisement video reasoning. On the AdsQA benchmark, it achieves 25.0% strict / 51.5% relaxed accuracy, outperforming open-source multimodal baselines of similar size. AdsQA contains 1,544 videos, 10,962 clips, and 22.7 hours of content across five high-level tasks.
