opendatalab/ChartVerse-Coder
ChartVerse-Coder by opendatalab is a 7.6 billion parameter, complexity-aware chart code generator built on Qwen2.5-Coder-7B-Instruct, designed to autonomously synthesize diverse, high-complexity chart codes from scratch. Unlike template-based methods, it uses high-temperature sampling and Rollout Posterior Entropy (RPE) filtering to explore long-tail chart distributions, producing realistic charts with advanced structural complexity. Its primary use case is generating a wide variety of complex visualizations, including 3D plots, hierarchical structures, and multi-subplot layouts, for data analysis and presentation.
Loading preview...
ChartVerse-Coder: Autonomous, Complexity-Aware Chart Code Generation
ChartVerse-Coder, developed by opendatalab as part of the ChartVerse project, is a 7.6 billion parameter model built on Qwen2.5-Coder-7B-Instruct. It specializes in autonomously generating diverse and complex chart codes from scratch, moving beyond traditional template-based or seed-conditioned approaches. The model leverages high-temperature sampling to explore a broad spectrum of chart distributions, producing highly varied and structurally intricate visualizations.
Key Capabilities
- Autonomous Synthesis: Generates diverse chart codes without relying on templates or initial seed charts.
- Complexity-Aware: Utilizes Rollout Posterior Entropy (RPE)-guided filtering during training to master high-complexity visualizations, ensuring generated charts are intricate and meaningful.
- High Diversity: Capable of producing a wide array of chart types, including 3D plots, hierarchical structures, multi-subplot layouts, and specialized charts like Sankey diagrams.
- Iterative Self-Enhancement: Employs a unique training pipeline involving generation, filtering (based on execution validity, RPE, and similarity), and retraining to progressively improve code quality and complexity.
Good for
- Generating complex data visualizations: Ideal for users needing to create intricate charts like 3D plots, treemaps, or multi-subplot dashboards.
- Exploring diverse chart types: Useful for discovering novel or less common visualization styles beyond standard templates.
- Automating chart creation: Developers can integrate ChartVerse-Coder to programmatically generate Python visualization code for various data analysis and reporting needs.