Add dataset card for PowerCLIP pre-processed CC12M

README.md

@@ -0,0 +1,38 @@
+---
+task_categories:
+- zero-shot-image-classification
+---
+
+# PowerCLIP: Powerset Alignment for Contrastive Pre-Training
+
+This repository contains the pre-processed CC12M dataset used in the paper [PowerCLIP: Powerset Alignment for Contrastive Pre-Training](https://huggingface.co/papers/2511.23170).
+
+PowerCLIP is a contrastive pre-training framework that enhances standard CLIP by optimizing region-to-phrase alignments. This dataset augments the standard CC12M image-text pairs with two additional semantic components:
+1. **SAM region extraction**: Segment Anything Model (SAM) generated semantic regions per image.
+2. **Parse-tree phrase extraction**: spaCy-based constituency parsing extracted phrases from captions.
+
+## Resources
+- **Paper:** [PowerCLIP: Powerset Alignment for Contrastive Pre-Training](https://huggingface.co/papers/2511.23170)
+- **GitHub Repository:** [KMasaki0210/PowerCLIP](https://github.com/KMasaki0210/PowerCLIP)
+
+## Dataset Structure
+The dataset is provided in WebDataset format. Each sample in the tar archives contains the following files:
+
+```
+{key}.jpg           # Image
+{key}.txt           # Caption
+{key}.json          # Metadata
+{key}.njson         # Parse-tree phrase indices (CSR format)
+{key}.samlens.npy   # SAM region lengths (CSR format)
+{key}.samcat.npy    # SAM region token indices (CSR format)
+```
+
+## Citation
+```bibtex
+@article{masaki2025powerclip,
+  title={PowerCLIP: Powerset Alignment for Contrastive Pre-Training},
+  author={Masaki, Kobayashi and others},
+  journal={arXiv preprint arXiv:2511.23170},
+  year={2025}
+}
+```