CewEhao/OPD-Aha-9B
OPD-Aha-9B by CewEhao is a 9 billion parameter multimodal model built upon Qwen3.5-9B, designed for enhanced fine-grained visual perception and multimodal mathematical reasoning. It utilizes an on-policy self-distillation framework with a frozen visual teacher and counterfactual visual inputs to focus learning on evidence that alters teacher distribution. This approach makes it particularly effective for tasks requiring detailed visual understanding and complex mathematical problem-solving within a 32768 token context.
Loading preview...
Overview
OPD-Aha-9B is a 9 billion parameter multimodal model developed by CewEhao, based on the Qwen3.5-9B architecture. Its core innovation lies in an on-policy self-distillation framework designed to significantly improve performance in two key areas: fine-grained visual perception and multimodal mathematical reasoning. The model learns by focusing on visual evidence that causes a change in its 'teacher' model's distribution, using a frozen visual teacher and counterfactual visual inputs.
Key Capabilities
- Enhanced Fine-Grained Visual Perception: Specialized training allows for more detailed and accurate understanding of visual information.
- Multimodal Mathematical Reasoning: Excels at solving mathematical problems that involve both textual and visual components.
- On-Policy Self-Distillation: Leverages a unique training method to refine its learning process, concentrating on impactful visual cues.
Good For
- Applications requiring precise visual analysis.
- Tasks involving complex mathematical problem-solving with visual context.
- Research and development in multimodal AI, particularly in distillation techniques for visual and mathematical understanding.
Serving instructions and training/evaluation code are available in the associated GitHub repository, which also includes evaluation scripts for benchmarks like V*Bench, HR-Bench, MathVista, and MathVerse.