Praha-Labs/Gemma-4-12B-TikZ-LoRA
Praha-Labs/Gemma-4-12B-TikZ-LoRA is a 12 billion parameter LoRA fine-tuned model based on unsloth/gemma-4-12b-it, specifically designed for instruction-to-TikZ generation. This model excels at converting natural language instructions into TikZ code, making it highly specialized for generating vector graphics. It was trained using BF16 LoRA with a focus on TikZ token loss, processing over 115 million input tokens.
Loading preview...
Overview
Praha-Labs/Gemma-4-12B-TikZ-LoRA is a 12 billion parameter LoRA (Low-Rank Adaptation) fine-tuned model, built upon the unsloth/gemma-4-12b-it base model. Its primary function is to translate natural language instructions into TikZ code, enabling the generation of vector graphics from text prompts. This model utilizes a BF16 LoRA training method, focusing specifically on optimizing for TikZ token generation.
Key Capabilities
- Instruction-to-TikZ Generation: Converts textual descriptions into executable TikZ code for creating diagrams and graphics.
- Specialized Fine-tuning: Trained with a focus on TikZ tokens, ensuring high relevance and accuracy for graphic generation tasks.
- Efficient LoRA Adaptation: Employs a LoRA adapter with 100,466,688 trainable parameters (3.8387% of total parameters) for efficient adaptation.
- Comprehensive Training Data: Trained on 94,675 examples, processing over 115 million input tokens, with loss calculated against reference TikZ implementations.
Good for
- Automated Diagram Generation: Ideal for developers and researchers needing to programmatically create TikZ diagrams from text.
- Technical Documentation: Useful for generating complex mathematical or scientific figures directly from descriptions.
- Educational Content Creation: Can assist in producing visual aids for teaching materials by converting instructions into TikZ.
- Rapid Prototyping of Graphics: Enables quick iteration and generation of vector graphics without manual TikZ coding.