kagakouko/Spatial-Interactor-Qwen2.5-VL-3B
The Spatial-Interactor Qwen2.5-VL-3B is a 3 billion parameter vision-language model developed by kagakouko, based on Qwen/Qwen2.5-VL-3B-Instruct, with a context length of 32768 tokens. It is specifically fine-tuned for spatial reasoning by learning local world-state and ego-motion transitions through supervised fine-tuning and On-Policy Distillation. This model excels at understanding and predicting interactions within the observable physical world, making it suitable for tasks requiring advanced spatial intelligence.
Loading preview...
Spatial-Interactor Qwen2.5-VL-3B Overview
This model, developed by kagakouko, is a 3 billion parameter vision-language model built upon the Qwen2.5-VL-3B-Instruct architecture. It is uniquely designed to enhance spatial reasoning capabilities by learning from interactions within the observable physical world. The model leverages a combination of supervised fine-tuning (SFT) and On-Policy Distillation (OPD) to integrate successive transitions over long trajectories, enabling it to understand and predict complex spatial dynamics.
Key Capabilities
- Advanced Spatial Reasoning: Learns local world-state and ego-motion transitions to infer spatial relationships and predict interactions.
- Interaction-based Learning: Utilizes a novel training approach involving On-Policy Distillation to integrate sequential observations.
- Vision-Language Integration: Processes both image/video and textual inputs, similar to its base model, without requiring additional trace or teacher branches during inference.
- Efficient Inference: Operates as a full-parameter BF16 checkpoint, providing robust performance with a 32768 token context length.
Good for
- Robotics and Embodied AI: Developing agents that need to understand and navigate physical environments.
- Video Understanding: Analyzing and predicting events in video streams that involve spatial interactions.
- Complex Scene Analysis: Tasks requiring deep comprehension of object relationships and movements within visual data.
- Research in Spatial Cognition: Exploring and advancing the field of AI with enhanced spatial intelligence.