RL-MIND/UHR-BAT2
UHR-BAT2 is an 8 billion parameter vision-language model developed by RL-MIND, fine-tuned from Qwen3-VL-8B-Instruct. It specializes in remote-sensing visual question answering, leveraging its base architecture with specific training on 10,000 GeoLLaVA examples. This model excels at interpreting satellite and aerial imagery for tasks like land use description and object identification, achieving notable accuracy on datasets such as RSHR-Bench (50.72%) and XLRS-Bench-lite (52.27%). Its primary use case is in remote sensing applications requiring detailed image analysis and question answering.
Loading preview...
UHR-BAT2: Remote-Sensing Visual Question Answering
UHR-BAT2 is an 8 billion parameter vision-language model developed by RL-MIND, specifically fine-tuned for remote-sensing visual question answering (VQA). It is built upon the Qwen3-VL-8B-Instruct architecture, retaining its native structure for seamless integration with standard Transformers libraries. The model was fine-tuned using a carefully selected 10,000-example manifest from the GeoLLaVA dataset, focusing on enhancing its ability to interpret and answer questions about remote sensing imagery.
Key Capabilities
- Specialized Remote Sensing VQA: Optimized for understanding and responding to queries related to satellite and aerial imagery.
- High Accuracy on RS Benchmarks: Achieves 50.72% on RSHR-Bench (All-QA option accuracy), 52.27% on XLRS-Bench-lite (Strict option-set accuracy), and 56.80% on MME-RealWorld Remote Sensing.
- Qwen3-VL Architecture: Leverages the robust Qwen3-VL-8B-Instruct base, ensuring compatibility and performance.
- Efficient Fine-tuning: Trained with a frozen vision encoder, focusing on the merger and LLM components over one epoch.
Good for
- Geospatial Analysis: Applications requiring automated interpretation of remote sensing data.
- Environmental Monitoring: Identifying land use, changes, and specific objects in aerial or satellite images.
- Defense and Intelligence: Analyzing reconnaissance imagery for specific features or activities.
- Academic Research: As a strong baseline or tool for further research in remote sensing AI.