Kwai-Keye/Thyme-SFT
Thyme-SFT is a 7 billion parameter supervised fine-tuned model developed by Kwai-Keye, designed to transcend traditional image processing by autonomously generating and executing diverse image processing and computational operations through executable code. With a 32768-token context length, it significantly enhances performance on high-resolution perception and complex reasoning tasks, leveraging a novel two-stage training strategy. This model is optimized for advanced visual reasoning and code-based image manipulation.
Loading preview...
Thyme-SFT: Think Beyond Images
Thyme-SFT is a 7 billion parameter supervised fine-tuned model from Kwai-Keye, designed to revolutionize how AI interacts with images. Unlike traditional models, Thyme-SFT focuses on autonomously generating and executing diverse image processing and computational operations via executable code. This approach significantly boosts performance in high-resolution perception and complex reasoning tasks.
Key Capabilities
- Code-based Image Manipulation: Generates and executes code for image processing and computational tasks.
- Enhanced Visual Reasoning: Improves performance on complex reasoning tasks involving high-resolution images.
- Advanced Perception: Excels in high-resolution image perception.
- Novel Training Strategy: Utilizes a two-stage training strategy, combining supervised fine-tuning with reinforcement learning, and empowered by the GRPO-ATS algorithm.
Good for
- Applications requiring advanced visual reasoning beyond simple image recognition.
- Tasks that benefit from AI-generated code for image processing.
- Research and development in autonomous image manipulation and computational vision.
- Scenarios demanding high-resolution perception and complex problem-solving in visual domains.