miniHui/Geo-R1
Geo-R1 is a 7 billion parameter vision-language model (VLM) developed by miniHui, specifically designed for geospatial reasoning tasks. It utilizes a unique post-training framework combining supervised fine-tuning on synthetic chain-of-thought exemplars and GRPO-based reinforcement learning with a cross-view pairing proxy. This architecture enables the model to connect visual cues with geographic priors, excelling at accurate prediction across various geospatial reasoning benchmarks with a 32768 token context length.
Loading preview...
Geo-R1: Geospatial Reasoning VLM
Geo-R1 is a 7 billion parameter vision-language model (VLM) developed by miniHui, specifically engineered to unlock advanced geospatial reasoning capabilities. This model introduces a novel post-training framework that significantly enhances a VLM's ability to interpret and reason about geographic information from visual inputs.
Key Capabilities
- Geospatial Reasoning: Excels at tasks requiring the interpretation of visual cues in conjunction with geographic knowledge.
- Cross-View Reinforcement Learning: Leverages a unique GRPO-based reinforcement learning approach on a weakly-supervised cross-view pairing proxy to improve reasoning.
- Chain-of-Thought Integration: Incorporates "thinking scaffolding" through supervised fine-tuning on synthetic chain-of-thought exemplars, fostering more robust reasoning processes.
- State-of-the-Art Performance: Achieves leading results across various geospatial reasoning benchmarks.
Good For
- Applications requiring precise geospatial understanding from visual data.
- Research and development in advanced VLM reasoning for geographic contexts.
- Tasks that benefit from connecting visual information with underlying geographic priors.
For more detailed information, refer to the accompanying paper: Geo-R1: Unlocking VLM Geospatial Reasoning with Cross-View Reinforcement Learning.