yongxianwei/Qwen2.5-VL-32B-Chart
yongxianwei/Qwen2.5-VL-32B-Chart is a 32 billion parameter vision-language model, fine-tuned from Qwen2.5-VL-32B-Instruct, specifically optimized for chart understanding and visual question answering on chart data. Developed by yongxianwei as part of the OptMerge MLLM merging benchmark, this model excels at interpreting complex visual information within charts. It is designed as a specialized expert for multimodal LLM merging, focusing on enhancing chart-related capabilities.
Loading preview...
Overview
yongxianwei/Qwen2.5-VL-32B-Chart is a 32 billion parameter vision-language model (VLM) specifically fine-tuned for chart understanding and visual question answering (VQA) on chart data. It is an expert model developed by yongxianwei as part of the OptMerge MLLM merging benchmark, designed to demonstrate how specialized capabilities can be integrated into a unified multimodal LLM.
Key Capabilities
- Specialized Chart Understanding: Optimized for interpreting various chart types and extracting information from them.
- Enhanced Chart VQA: Achieves strong performance on datasets like ChartQA, outperforming its base model.
- OptMerge Expert: Functions as one of five capability experts (VQA, Geometry, Chart, OCR, Grounding) that can be merged using the OptMerge method to create a more comprehensive multimodal model.
- Full Weights: Provided as full weights fine-tuned from Qwen/Qwen2.5-VL-32B-Instruct, requiring no PEFT adapter for direct loading.
Good For
- Chart-based Question Answering: Ideal for applications requiring precise information extraction and reasoning from visual charts.
- Research on Model Merging: Serves as a component for researchers exploring multimodal LLM merging techniques, particularly with the OptMerge framework.
- Benchmarking Chart Understanding: Useful for evaluating and comparing VLM performance on chart-specific tasks.
This model demonstrates improved performance on ChartQA (86.96%) compared to its base model (83.64%), as detailed in the OptMerge paper.