InternVL2_5-8B-MPO-AWQ
| Property | Value |
|---|---|
| Model Size | 8B parameters |
| License | MIT License |
| Architecture | ViT-MLP-LLM with InternViT-300M-448px-V2_5 + internlm2_5-7b-chat |
| Paper | arXiv:2411.10442 |
What is InternVL2_5-8B-MPO-AWQ?
InternVL2_5-8B-MPO-AWQ is an advanced multimodal large language model that combines sophisticated vision processing capabilities with powerful language understanding. It represents a significant advancement in the InternVL family, incorporating Mixed Preference Optimization (MPO) to enhance reasoning abilities and overall performance in visual-linguistic tasks.
Implementation Details
The model utilizes a "ViT-MLP-LLM" architecture, featuring an InternViT vision encoder connected to a language model through an MLP projector. It processes images using a 448×448 pixel dynamic resolution strategy and includes innovative features for handling multi-image and video data.
- Implements Mixed Preference Optimization (MPO) combining preference, quality, and generation loss
- Uses DPO as preference loss and BCO as quality loss
- Supports multi-image processing with dynamic resolution
- Features AWQ quantization for efficient deployment
Core Capabilities
- Strong performance across multiple benchmarks (82.4% on MMBench v1.1)
- Enhanced reasoning abilities through MPO training
- Efficient multi-image and video processing
- Superior OCR capabilities (88.3% on OCRBench)
- Advanced visual reasoning and description generation
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its Mixed Preference Optimization approach, which combines three different types of learning objectives to enhance reasoning capabilities and response quality. This, coupled with its efficient architecture and multi-modal processing abilities, makes it particularly effective for complex visual-linguistic tasks.
Q: What are the recommended use cases?
The model excels in visual description, reasoning tasks, OCR applications, and multi-image analysis. It's particularly suitable for applications requiring detailed image understanding, mathematical visual reasoning, and complex visual question-answering scenarios.
