nllg/TikZilla-3B-RL

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:3.1BQuant:BF16Ctx Length:32kPublished:Mar 9, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

TikZilla-3B-RL is a 3.1 billion parameter language model developed by nllg, specifically designed for generating TikZ/LaTeX figures from natural language descriptions. Based on TikZilla-3B, this model was fine-tuned using reinforcement learning (RL) on the DaTikZ-V4 dataset. It excels at translating textual figure requirements into complete and valid LaTeX code, making it highly specialized for scientific figure generation.

Loading preview...

TikZilla-3B-RL: Specialized TikZ/LaTeX Figure Generation

TikZilla-3B-RL is a 3.1 billion parameter language model developed by nllg, engineered for the precise task of generating TikZ/LaTeX figures from natural language prompts. This model builds upon the foundation of TikZilla-3B and has undergone reinforcement learning (RL) using the DaTikZ-V4 dataset, specifically optimizing its performance for scientific figure generation.

Key Capabilities

  • Natural Language to TikZ/LaTeX: Translates detailed textual descriptions into complete and valid LaTeX documents containing TikZ figures.
  • Scientific Figure Generation: Optimized for creating common scientific plots and diagrams, such as line plots with labeled axes, legends, and specific styling.
  • Reinforcement Learning Enhanced: Benefits from RL training on a specialized dataset, likely improving the quality and accuracy of generated TikZ code.

Good For

  • Automating Figure Creation: Researchers and academics can use this model to quickly generate complex scientific figures without manual TikZ coding.
  • Prototyping Visualizations: Rapidly create prototypes of figures based on high-level descriptions.
  • Integrating into Workflows: Can be integrated into LaTeX-based document generation pipelines to streamline figure production.