README.md

---
language:
- en
license: cc-by-nc-4.0
size_categories:
- 100K<n<1M
task_categories:
- image-text-to-text
pretty_name: EditReward-Data
tags:
- image-editing
- reward-modeling
- preference-data
dataset_info:
  features:
  - name: key
    dtype: string
  - name: unique_key
    dtype: string
  - name: pair_index
    dtype: int32
  - name: instruction
    dtype: string
  - name: source_img
    dtype: image
  - name: left_output_image
    dtype: image
  - name: right_output_image
    dtype: image
  - name: left_model
    dtype: string
  - name: right_model
    dtype: string
  - name: left_overall_score
    dtype: float32
  - name: left_instruction_following_score
    dtype: float32
  - name: left_quality_aesthetic_score
    dtype: float32
  - name: right_overall_score
    dtype: float32
  - name: right_instruction_following_score
    dtype: float32
  - name: right_quality_aesthetic_score
    dtype: float32
  - name: vote_type
    dtype: string
  - name: metadata_index
    dtype: int32
  - name: left_img_rel
    dtype: string
  - name: right_img_rel
    dtype: string
  splits:
  - name: train
    num_bytes: 338525893077.312
    num_examples: 170772
  download_size: 86733810500
  dataset_size: 338525893077.312
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# EditReward-Data

This repository contains **EditReward-Data**, a large-scale, high-fidelity human preference dataset for instruction-guided image editing. It was introduced in the paper [EditReward: A Human-Aligned Reward Model for Instruction-Guided Image Editing](https://huggingface.co/papers/2509.26346).

`EditReward-Data` comprises over 200K manually annotated preference pairs. These annotations were meticulously curated by trained experts following a rigorous and standardized protocol, ensuring high alignment with considered human judgment and minimizing label noise. The dataset covers a diverse range of edits produced by seven state-of-the-art models across twelve distinct sources. It serves as crucial training data for reward models like EditReward, designed to score instruction-guided image edits.

- **Paper:** [EditReward: A Human-Aligned Reward Model for Instruction-Guided Image Editing](https://huggingface.co/papers/2509.26346)
- **Project Page:** [https://tiger-ai-lab.github.io/EditReward](https://tiger-ai-lab.github.io/EditReward)
- **Code Repository:** [https://github.com/TIGER-AI-Lab/EditReward](https://github.com/TIGER-AI-Lab/EditReward)

<p align="center">
  <img src="https://github.com/TIGER-AI-Lab/EditReward/blob/main/assets/pipeline.png?raw=true" alt="EditReward Pipeline" width="900"/>
</p>

## Dataset Overview

EditReward-Data is designed to enable the training of reward models that can score instruction-guided image edits. The dataset facilitates assessing and improving the alignment of image editing models with human preferences. The dataset statistics are shown below:

<p align="left">
  <img src="https://github.com/TIGER-AI-Lab/EditReward/blob/main/assets/dataset_stat.png?raw=true" alt="Dataset Statistics" width="900"/>
</p>

## Sample Usage

To download the `EditReward-Data` dataset to your local machine, use the `huggingface-cli` command:

```bash
huggingface-cli download --repo-type dataset TIGER-Lab/EditReward-Data --local-dir /your-local-dataset-path
```

## Citation

Please kindly cite our paper if you use our code, data, models, or results:

```bibtex
@article{wu2025editreward,
  title={EditReward: A Human-Aligned Reward Model for Instruction-Guided Image Editing},
  author={Wu, Keming and Jiang, Sicong and Ku, Max and Nie, Ping and Liu, Minghao and Chen, Wenhu},
  journal={arXiv preprint arXiv:2509.26346},
  year={2025}
}
```