opendatalab/ChartVerse-8B
ChartVerse-8B is an 8-billion parameter Vision Language Model (VLM) developed by opendatalab, specifically optimized for chart reasoning tasks. It achieves a 64.1% average score across 6 challenging chart benchmarks, outperforming its 30B parameter teacher model and approaching the performance of a 32B model. This VLM leverages high-quality synthetic data and a 32,768-token context length to excel in complex chart analysis and question answering.
Loading preview...
ChartVerse-8B: A High-Performance VLM for Chart Reasoning
ChartVerse-8B, developed by opendatalab, is an 8-billion parameter Vision Language Model (VLM) specifically designed for advanced chart reasoning. It demonstrates exceptional performance, achieving a 64.1% average score across six challenging chart benchmarks. Notably, this model surpasses its 30-billion parameter teacher model (Qwen3-VL-30B-A3B-Thinking) and approaches the performance of a 32-billion parameter model (Qwen3-VL-32B-Thinking), showcasing the effectiveness of its training methodology.
Key Capabilities
- Superior Chart Reasoning: Achieves a 64.1% average score on chart benchmarks, indicating strong analytical capabilities for various chart types.
- Efficient Performance: Outperforms larger teacher models with significantly fewer parameters, demonstrating high efficiency.
- Advanced Training: Utilizes a unique training regimen involving Supervised Fine-Tuning (SFT) on 600K high-complexity charts with CoT reasoning and Reinforcement Learning (RL) on 40K difficult samples.
- High Context Length: Supports a context length of 32,768 tokens, enabling comprehensive analysis of complex charts and associated queries.
Good for
- Automated Chart Analysis: Ideal for applications requiring automated extraction of insights and answers from visual charts.
- Data Interpretation Systems: Suitable for integrating into systems that need to interpret and reason about data presented in graphical formats.
- Research in VLM Distillation: Provides a strong example of how smaller student models can exceed larger teacher models through effective synthetic data generation and training.