nllg/EdiTikZ-4B

EdiTikZ-4B: Multimodal TikZ Figure Editing and Reconstruction

EdiTikZ-4B is a specialized 4.5 billion parameter multimodal language model, built upon the Qwen3.5-4B architecture. Developed by nllg, its core function is to process scientific figures and generate or modify their corresponding TikZ/LaTeX code based on user instructions.

Key Capabilities

• Instruction-Guided Editing: Users can provide natural language instructions to modify existing scientific figures. The model interprets these instructions and outputs updated TikZ/LaTeX code reflecting the requested changes.
• Image-to-TikZ Reconstruction: EdiTikZ-4B can reconstruct scientific figures from images into editable TikZ/LaTeX code, enabling programmatic manipulation of visual content.
• Specialized Training: The model underwent multi-task supervised fine-tuning (SFT) using the DaEdiTikZ and DaTikZ-V4 datasets, specifically tailored for scientific figure editing and reconstruction tasks.
• TikZ/LaTeX Output: It generates valid LaTeX code, ensuring compatibility with standard LaTeX compilation workflows for scientific documents.

Good For

• Automated Figure Generation: Creating TikZ code from visual representations of scientific figures.
• Programmatic Figure Editing: Applying precise, instruction-based modifications to TikZ figures without manual code adjustments.
• Scientific Document Preparation: Streamlining the process of incorporating and customizing complex scientific diagrams in LaTeX documents.