Add dataset card, link to paper and code

README.md

@@ -1,3 +1,41 @@
 ---
 license: cc-by-4.0
+task_categories:
+- text-to-image
 ---
+
+# RealAlign Dataset
+
+[**Project Page**](https://cwyxx.github.io/RealAlign) | [**Paper**](https://huggingface.co/papers/2605.19839) | [**Code**](https://github.com/Cwyxx/RealAlign)
+
+This dataset is part of the **RealAlign** project, introduced in the paper "[When Preference Labels Fall Short: Aligning Diffusion Models from Real Data](https://huggingface.co/papers/2605.19839)".
+
+## Dataset Summary
+
+RealAlign investigates whether real data can serve as an alternative source of supervision for preference alignment in diffusion models. Instead of relying on manually annotated preference pairs from model-generated images, this dataset treats real images as reference points and constructs preference signals by contrasting them with generated or perturbed samples.
+
+The repository contains curated `(real, fake)` preference pairs derived from:
+- **HPDv3**
+- **Pick-a-Pic v2**
+- **Civitai-top**
+
+## Data Curation Pipeline
+
+The dataset was built using a four-stage curation pipeline:
+1. **Extract**: Initial data collection from source distributions.
+2. **Construct Pairs**: Creating pairings between real images and model-generated or perturbed samples.
+3. **Score**: Applying automated scoring to evaluate the pairs.
+4. **Filter**: Final selection of high-quality preference signals to produce the training CSV.
+
+These curated pairs are used to align diffusion models like Stable Diffusion 1.5 and Stable Diffusion 3.5 Medium through a two-stage training process involving Diffusion-DRO (Inverse RL) and Diffusion-DPO.
+
+## Citation
+
+```bibtex
+@article{realalign2024,
+  title={When Preference Labels Fall Short: Aligning Diffusion Models from Real Data},
+  author={Zhou, Yifan and others},
+  journal={arXiv preprint arXiv:2605.19839},
+  year={2024}
+}
+```