| Property | Value |
|---|---|
| Parameter Count | 2 Billion |
| Model Type | Vision-Language Model |
| Architecture | Qwen2-VL-2B-Instruct |
| Languages | English, Chinese |
| Repository | EvolvingLMMs-Lab/open-r1-multimodal |
| Training Data | lmms-lab/multimodal-open-r1-8k-verified |
What is Qwen2-VL-2B-GRPO-8k?
Qwen2-VL-2B-GRPO-8k is an advanced multimodal model that builds upon the Qwen2-VL-2B-Instruct architecture, fine-tuned on a carefully curated dataset of 8,000 samples using GRPO methodology. It's designed to handle both vision and language tasks effectively, supporting both English and Chinese languages.
Implementation Details
The model utilizes the Transformers library and can be easily implemented using Qwen2VLForConditionalGeneration. It features flexible visual token processing capabilities, with customizable pixel ranges for balancing speed and memory usage. The model employs a structured system prompt approach for generating responses, incorporating both reasoning and answer components.
- Supports dynamic visual token processing (4-16384 tokens per image)
- Implements bfloat16 precision for efficient computation
- Features a chat template system for structured responses
- Includes built-in reasoning capabilities with think/answer frameworks
Core Capabilities
- Multimodal understanding and generation
- Bilingual support (English and Chinese)
- Flexible image processing with customizable resolution
- Structured reasoning and response generation
- Efficient memory management with adjustable visual token ranges
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its combination of the Qwen2-VL architecture with GRPO training on a carefully curated 8k dataset, offering balanced performance and efficiency in vision-language tasks while maintaining bilingual capabilities.
Q: What are the recommended use cases?
The model is ideal for applications requiring image description, visual question answering, and multimodal reasoning tasks in both English and Chinese languages. It's particularly suitable for scenarios where efficient processing of visual information needs to be balanced with accurate language generation.