Improve dataset card with task category, paper link, and details

This PR improves the dataset card by:
- Adding the `image-classification` task category to the metadata.
- Including a direct link to the paper.
- Expanding the description with more details from the GitHub README, clarifying the dataset's contents and preprocessing steps.

README.md

@@ -1,5 +1,13 @@
 ---
 license: mit
+task_categories:
+- image-classification
 ---
 
-The official dataset for arxiv.org/abs/2506.03131.
+This repository contains the preprocessed ImageNet1K dataset used for training the Native-resolution diffusion Transformer (NiT) model, as presented in [Native-Resolution Image Synthesis](https://huggingface.co/papers/2506.03131). The dataset is available in preprocessed latent format for efficient training.
+
+**Dataset Contents:** The dataset includes preprocessed latent representations of images from ImageNet1K at various resolutions (256x256, 512x512, and native resolutions). The preprocessing steps involve converting images into latent space using a diffusion model. Pre-packed sampler metadata is also provided, optimizing data loading for efficient training of the NiT model. Specific details on preprocessing and packing can be found in the original repository's documentation [here](https://github.com/WZDTHU/NiT).
+
+**Project Page:** [https://wzdthu.github.io/NiT](https://wzdthu.github.io/NiT)
+
+**Note:** To use this dataset effectively, please refer to the original repository's instructions on dataset download and usage. The dataset is optimized for training the NiT model and might require specific tools and libraries.