loss-boss/tikz-dataset

Mein Dataset

Dataset Description

This dataset is a cleaned and extended collection of image–TikZ code pairs for training and evaluating image-to-TikZ models.

It contains processed data derived from:

• DaTikZ-V4
• GeoTikz-Base

Additional samples were collected using the scripts provided by the creators of DaTikZ:

• DaTikZ repository

Data Cleaning

A major issue in image-to-TikZ datasets is that some LaTeX files generate multi-page PDFs, while the stored image contains only the first page. In these cases, the image and target code do not represent the same content.

We compiled the TikZ and LaTeX sources and removed all samples that produced more than one PDF page. This improves image–code alignment for both training and benchmarking.

For GeoTikz-Base, some figures were rendered incorrectly because required LaTeX packages or TikZ libraries were missing. These samples were rendered again using a more complete LaTeX environment.

We also removed Markdown code fences from the GeoTikz response field. The cleaned field therefore contains only the raw LaTeX or TikZ source code.

Dataset Splits

• datikz_v4: cleaned samples from DaTikZ-V4
• geotikz_bridge_base: cleaned and re-rendered samples from GeoTikz-Base
• our_dataset_train: additionally collected training samples
• our_dataset_benchmark: held-out samples for model evaluation

Intended Use

The dataset is intended for:

• Image-to-TikZ model training
• Multimodal fine-tuning
• Model benchmarking

Splits

• datikz_v4
• geotikz_bridge_base
• our_dataset_benchmark
• our_dataset_train