TsinghuaC3I/Qwen2.5-7B-VL-ReAd-R
TsinghuaC3I/Qwen2.5-7B-VL-ReAd-R is a 7 billion parameter video understanding model based on Qwen2.5-VL-7B, developed by TsinghuaC3I. It is fine-tuned using GRPO reinforcement learning specifically for advertisement video reasoning. This model excels at analyzing advertisement content, achieving 25.0% strict and 51.5% relaxed accuracy on the AdsQA benchmark, outperforming other open-source multimodal baselines of similar size.
Loading preview...
Qwen2.5-7B-VL-ReAd-R: Advertisement Video Reasoning Model
TsinghuaC3I/Qwen2.5-7B-VL-ReAd-R is a specialized video understanding model built upon the Qwen2.5-VL-7B architecture, developed by TsinghuaC3I. This 7 billion parameter model is uniquely fine-tuned using GRPO (Generalized Reinforcement Learning with Policy Optimization) to excel in the domain of advertisement video reasoning. Its primary focus is on understanding and analyzing the content within advertising videos.
Key Capabilities
- Advertisement Video Reasoning: Specifically optimized for tasks involving the analysis and interpretation of advertisement video content.
- Reinforcement Learning Fine-tuning: Utilizes GRPO for enhanced performance in its target domain.
- Competitive Performance: Achieves 25.0% strict accuracy and 51.5% relaxed accuracy on the AdsQA benchmark, demonstrating superior performance compared to other open-source multimodal models of comparable size.
Good For
- Analyzing Advertisement Content: Ideal for applications requiring detailed understanding of advertising videos.
- Video Understanding Benchmarking: Suitable for research and development in video reasoning, particularly on datasets like AdsQA which features 1,544 videos and 22.7 hours of content across various tasks.
- Developing Specialized Multimodal AI: A strong baseline for projects focused on niche video analysis tasks where general-purpose models may fall short.