gemma4-4b-tikz

gemma4-4b-tikz is a TikZ-focused fine-tune of Gemma 4, trained to improve generation of compilable TikZ/LaTeX figures from scientific captions and multimodal inputs.

Model Description

• Developed by: Michele Banfi
• Base model: unsloth/gemma-4-E4B-it
• Method: QLoRA (4-bit) + SFT via Unsloth
• Dataset: nllg/DaTikZ-V4
• License: Gemma Terms of Use

Intended Use

• Generate TikZ/LaTeX code from scientific figure descriptions
• Improve robustness of TikZ syntax for compilable outputs
• Support multimodal TikZ generation workflows

Training Setup (Current)

• 1 epoch (configurable in tikz/config.yaml)
• Mixed task training:
  • caption_to_code
  • image_to_description
  • image_to_code

Evaluation

Primary metric is compilation success rate on nllg/datikz-v3 test split, with optional visual similarity metrics (SSIM/LPIPS) for successful compilations.

Model Sources

• Repository: https://github.com/michelebanfi/finetune-gemma-4
• Sub-project docs: tikz/

Citation

@misc{banfi2026gemma4tikz,
  title={Gemma-4 TikZ Fine-tuning},
  author={Banfi, Michele},
  year={2026},
  howpublished={\url{https://github.com/michelebanfi/finetune-gemma-4}}
}