OpenMOSS-Team/Game-RL-Qwen2.5-VL-7B
Game-RL-Qwen2.5-VL-7B is a 7 billion parameter vision-language model developed by OpenMOSS-Team, fine-tuned from Qwen2.5-VL-7B using the GRPO strategy on the GameQA-5K dataset. This model specializes in multimodal reasoning, particularly for game-code-driven scenarios, and demonstrates enhanced out-of-domain generalization compared to models trained on other datasets. Its primary use case is advanced visual and logical reasoning within gaming contexts, leveraging game code for data synthesis.
Loading preview...
Game-RL-Qwen2.5-VL-7B: Enhanced Game-Code-Driven Multimodal Reasoning
This model, Game-RL-Qwen2.5-VL-7B, is a 7 billion parameter vision-language model (VLM) fine-tuned from Qwen2.5-VL-7B. It was trained using the GRPO strategy exclusively on the GameQA-5K dataset, which is derived from the larger GameQA-140K dataset. The GameQA dataset is unique as it is synthesized using a novel Code2Logic approach that leverages game code to generate multimodal reasoning data.
Key Capabilities
- Game-Code-Driven Reasoning: First VLM to be trained using game code to synthesize multimodal reasoning data, enhancing its understanding of game logic and scenarios.
- Enhanced Out-of-Domain Generalization: Demonstrates significantly improved generalization capabilities when trained with GRPO on GameQA, outperforming models trained on other datasets like MAVIS or multimodal-open-r1-8k-verified.
- Multimodal Understanding: Integrates visual and linguistic information for complex reasoning tasks, particularly within gaming contexts.
Good For
- Game-related AI applications: Ideal for tasks requiring deep understanding and reasoning within video game environments, such as game AI, interactive storytelling, or automated gameplay analysis.
- Research in VLM generalization: Useful for researchers exploring methods to improve VLM performance on novel or out-of-domain tasks through specialized data synthesis and training strategies.
- Developing robust multimodal agents: Provides a strong foundation for building agents that can interpret and act upon complex visual and logical information in dynamic, game-like settings.