Add dataset README

README.md

@@ -0,0 +1,48 @@
+# Flickr30k CLIP-Preprocessed Dataset
+
+This dataset contains the Flickr30k dataset preprocessed with CLIP ViT-Large-Patch14 image processor.
+
+## Files
+
+- `img_data.parquet`: Preprocessed images as flattened numpy arrays (shape: [3, 224, 224] -> flattened)
+- `train_captions.parquet`: Training split captions with image_id mapping
+- `val_captions.parquet`: Validation split captions with image_id mapping
+
+## Usage
+
+```python
+import pandas as pd
+import torch
+import numpy as np
+
+# Load the data
+images_df = pd.read_parquet('img_data.parquet')
+train_captions_df = pd.read_parquet('train_captions.parquet')
+val_captions_df = pd.read_parquet('val_captions.parquet')
+
+# Access an image
+image_id = 0
+flat_image = images_df.iloc[image_id]['image']
+image_tensor = torch.from_numpy(flat_image.reshape(3, 224, 224))
+
+# Access captions for that image
+captions_for_image = train_captions_df[train_captions_df['image_id'] == image_id]['caption'].tolist()
+```
+
+## Original Dataset
+
+Original dataset: [nlphuji/flickr30k](https://huggingface.co/datasets/nlphuji/flickr30k)
+
+## Preprocessing
+
+Images were processed using the CLIP ViT-Large-Patch14 image processor:
+- Resized to 224x224
+- CLIP normalization applied
+- Converted to tensors and flattened for storage efficiency
+
+## Dataset Statistics
+
+- Total images: Check `img_data.parquet` length
+- Train captions: Check `train_captions.parquet` length  
+- Validation captions: Check `val_captions.parquet` length
+- Train/Validation split: 90/10