---
license: apache-2.0
pipeline_tag: image-text-to-text
task_categories:
- image-to-text
- text-generation
language:
- en
size_categories:
- n<1K
tags:
- tikz
- latex
- code-generation
- scientific-figures
- image-editing
- multimodal
- benchmark
---

# Dataset Card for DaEdiTikZ-Bench

DaEdiTikZ-Bench is a human-curated benchmark for instruction-guided scientific figure editing in TikZ/LaTeX that has been used to evaluate our **EdiTikZ** models.

It contains 345 paired figures with complete TikZ programs, rendered PNG images, and bidirectional edit instructions (690 samples).

## Dataset fields

Each row contains:

- `sample_id`: benchmark sample identifier
- `similarity_score`: cosine similarity of the paired rendered figures (between 0.95 and 0.9999)
- `tikz_code_i`: complete TikZ/LaTeX source for the first figure
- `image_i`: rendered 448×448 PNG for the first figure
- `tikz_code_j`: complete TikZ/LaTeX source for the second figure
- `image_j`: rendered 448×448 PNG for the second figure
- `edit_instruction_forward`: detailed changes describing `i → j`
- `edit_instruction_backward`: detailed changes describing `j → i`

## Installation

```bash
pip install datasets
```

## Usage

```python
from datasets import load_dataset

dataset_id = "nllg/DaEdiTikZ-Bench"

ds = load_dataset(dataset_id, split="test")

pair = ds[0]

# Forward evaluation: image_i + forward instruction -> tikz_code_j
source_image = pair["image_i"]
instruction = pair["edit_instruction_forward"]
reference_tikz = pair["tikz_code_j"]
reference_image = sample["image_j"]

# Backward evaluation: image_j + backward instruction -> tikz_code_i
reverse_image = pair["image_j"]
reverse_instruction = pair["edit_instruction_backward"]
reverse_reference_tikz = pair["tikz_code_i"]
reverse_reference_image = pair["image_i"]
```