TIGER-Lab/VisCoder2-3B
VisCoder2-3B by TIGER-Lab is a 3.1 billion parameter multi-language visualization coding model, based on Qwen2.5-Coder-3B-Instruct. It is specifically trained for executable code generation, rendering, and iterative self-debugging across 12 programming languages. This model excels at generating semantically consistent visual outputs from natural language instructions, addressing challenges in multi-language visualization tasks.
Loading preview...
VisCoder2-3B: Multi-Language Visualization Coding Model
VisCoder2-3B, developed by TIGER-Lab, is a lightweight 3.1 billion parameter model designed for multi-language visualization coding. Built upon the Qwen2.5-Coder-3B-Instruct base model, it specializes in generating executable code, rendering visualizations, and performing iterative self-debugging.
Key Capabilities
- Multi-Language Support: Trained on the VisCode-Multi-679K dataset, it supports executable visualization tasks across 12 programming languages.
- Semantic Consistency: Addresses the challenge of generating code that not only executes successfully but also produces semantically consistent visual outputs aligned with natural language instructions.
- Iterative Self-Debugging: Demonstrates notable improvements in performance under multi-round self-debugging settings, as evaluated on the VisPlotBench.
- Performance: Achieves consistent performance across multiple languages on the VisPlotBench, which includes 888 executable visualization tasks spanning 8 languages.
Training Details
VisCoder2-3B was developed using full-parameter supervised fine-tuning (SFT) with the ms-swift framework, leveraging the extensive VisCode-Multi-679K instruction-tuning dataset.
Good For
- Developers and researchers working on automated visualization code generation.
- Applications requiring executable code generation from natural language for various programming languages.
- Tasks benefiting from iterative debugging capabilities in visualization contexts.