StevenHH2000/Fine-R1-7B
Fine-R1-7B by StevenHH2000 is a 7 billion parameter Multi-modal Large Language Model (MLLM) with a 32768 token context length, specifically designed for Fine-Grained Visual Recognition (FGVR). It utilizes Chain-of-Thought Supervised Fine-tuning (CoT-SFT) and Triplet Augmented Policy Optimization (TAPO) to excel at distinguishing highly similar visual sub-categories. This model bridges the gap between generative MLLMs and specialized discriminative models, outperforming general reasoning MLLMs in FGVR tasks.
Loading preview...
Fine-R1: Multi-modal LLM for Fine-Grained Visual Recognition
Fine-R1 is a 7 billion parameter Multi-modal Large Language Model (MLLM) developed by StevenHH2000, specifically engineered to excel in Fine-Grained Visual Recognition (FGVR). Unlike general MLLMs that often struggle with subtle visual distinctions, Fine-R1 is optimized to differentiate between highly similar sub-categories, bridging the performance gap with specialized discriminative models like CLIP.
Key Capabilities
- Enhanced FGVR Performance: Designed to identify minute visual differences, outperforming many general reasoning MLLMs and contrastive models in FGVR tasks.
- Chain-of-Thought Reasoning: Employs Chain-of-Thought Supervised Fine-tuning (CoT-SFT) on high-quality FGVR CoT datasets, enabling the model to perform detailed visual analysis and comparative reasoning before prediction.
- Robust Sub-category Distinction: Utilizes Triplet Augmented Policy Optimization (TAPO), including Intra-class Augmentation for visual variance and Inter-class Augmentation for maximizing distinction between similar sub-categories.
- Efficient Learning: Demonstrates strong performance in identifying both seen and unseen sub-categories with only 4-shot training.
Good For
- Applications requiring precise visual classification of highly similar objects or species.
- Research and development in advanced multi-modal AI, particularly in visual recognition and reasoning.
- Scenarios where traditional MLLMs lack the necessary granularity for visual discrimination.