internlm/Spatial-SSRL-7B
Spatial-SSRL-7B is a 7 billion parameter large vision-language model developed by InternLM, built upon Qwen2.5-VL-7B with a 32K context length. It is specifically optimized for spatial understanding tasks in both image and video settings through a lightweight, self-supervised reinforcement learning paradigm called Spatial-SSRL. This model excels at spatial intelligence while retaining the general visual capabilities of its base model, making it suitable for applications requiring precise spatial reasoning without extensive manual annotations.
Loading preview...
Spatial-SSRL-7B: Enhanced Spatial Understanding LVLM
Spatial-SSRL-7B is a 7 billion parameter large vision-language model (LVLM) developed by InternLM, specifically designed to enhance spatial understanding. Built on the Qwen2.5-VL-7B base model, it incorporates a novel self-supervised reinforcement learning (SSRL) paradigm to improve spatial intelligence across various benchmarks.
Key Capabilities
- Superior Spatial Understanding: Demonstrates stronger spatial intelligence in both image and video contexts, outperforming base models on seven spatial understanding benchmarks.
- Lightweight and Tool-Free: The Spatial-SSRL framework is designed to be lightweight and does not rely on external tools or extensive annotations, making it cost-effective and easily extensible.
- Highly Scalable Training: Utilizes ordinary raw RGB and RGB-D images for data curation, eliminating the need for richly-annotated public datasets or manual labels, which allows for high scalability.
- Cost-Effective Optimization: Avoids human labeling or API calls for general LVLMs throughout the training pipeline, significantly reducing costs.
- Naturally Verifiable: Intrinsic supervisory signals from pretext objectives ensure natural verifiability, aligning well with the RLVR paradigm.
Good For
- Applications requiring advanced spatial reasoning from visual data.
- Research and development in self-supervised learning for vision-language models.
- Scenarios where data annotation is costly or impractical, leveraging its scalable and cost-effective training approach.
- Developers looking for a vision-language model with strong spatial intelligence that preserves general visual capabilities.