HelloKKMe/GTA1-7B
HelloKKMe/GTA1-7B is a 7 billion parameter open-source GUI grounding model developed by HelloKKMe, specifically designed for precise UI element localization. Utilizing Reinforcement Learning (RL) with GRPO, it directly incentivizes actionable and grounded responses, outperforming other open-source models on ScreenSpot-V2, ScreenSpotPro, and OSWORLD-G benchmarks. This model excels at identifying exact coordinates of UI elements from images, making it ideal for automated GUI interaction and testing.
Loading preview...
Overview
HelloKKMe/GTA1-7B is a 7 billion parameter open-source model developed by HelloKKMe, specialized in GUI grounding. It leverages Reinforcement Learning (RL) with a method called GRPO (Grounded Reinforcement Policy Optimization) to achieve precise UI element localization. Unlike models that rely on verbose Chain-of-Thought (CoT) reasoning, GTA1-7B directly rewards successful clicks and actionable responses, ensuring better objective alignment for grounding tasks. This approach is based on findings from HelloKKMe's research on grounding.
Key Capabilities
- State-of-the-art GUI Grounding: Consistently achieves the best results among open-source model families on challenging datasets like ScreenSpot-V2, ScreenSpotPro, and OSWORLD-G.
- Reinforcement Learning (GRPO): Utilizes GRPO for objective alignment, directly incentivizing grounded and actionable responses rather than lengthy textual reasoning.
- Precise Coordinate Output: Designed to output exact (x,y) coordinates for specified UI elements, including the center point for elements with area.
Good for
- Automated GUI Interaction: Ideal for applications requiring precise identification and interaction with graphical user interface elements.
- UI Testing and Automation: Can be used to automate testing workflows by accurately locating and interacting with UI components.
- Robotics and Agent Control: Suitable for scenarios where an agent needs to understand and interact with visual interfaces based on natural language instructions.