---
license: apache-2.0
pipeline_tag: image-text-to-text
task_categories:
- image-to-text
- text-generation
language:
- en
size_categories:
- 100K<n<1M
tags:
- tikz
- latex
- code-generation
- scientific-figures
- image-editing
- multimodal
---

# Dataset Card for DaEdiTikZ

DaEdiTikZ is a dataset of realistic scientific-figure editing trajectories for TikZ/LaTeX used to train **EdiTikZ-4B**, **EdiTikZ-4B-RL**, **EdiTikZ-9B**, and **EdiTikZ-9B-RL**.

Each record pairs two source TikZ programs/rendered figures mined from **ArXiv**, **GitHub**, and **TeXStackExchange** revision trajectories, and natural-language edit instructions generated by **Qwen3.6-27B** in both directions. It supports the task of generating an edited TikZ program from a source figure and an edit instruction.

The release contains 106,358 redistributable figure pairs. Each pair provides two directional training instances. Image `i` to image `j` with the forward instruction, and image `j` to image `i` with the backward instruction, resulting in 212,716 samples.

## Dataset fields

Each row contains one bidirectional figure pair:

- `group_id`: source-specific revision group identifier
- `source`: origin category (`arxiv`, `github`, or `tex`)
- `similarity_score`: cosine similarity of the paired rendered figures (between 0.95 and 0.9999)
- `file_id_i`, `tikz_code_i`, `image_i`: identifier, full source, and rendered PNG for the first figure
- `file_id_j`, `tikz_code_j`, `image_j`: identifier, full source, and rendered 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"

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

pair = ds[0]

print(pair["group_id"], pair["source"], pair["similarity_score"])

# Forward editing task: image_i + forward instruction -> tikz_code_j
source_image = pair["image_i"]
instruction = pair["edit_instruction_forward"]
target_tikz = pair["tikz_code_j"]

# Reverse direction: image_j + backward instruction -> tikz_code_i
reverse_image = pair["image_j"]
reverse_instruction = pair["edit_instruction_backward"]
reverse_target_tikz = pair["tikz_code_i"]
```