michelinolinolino/gemma4-4b-tikz

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:
• —captiontocode
• —imagetodescription
• —imagetocode

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}}
}