armeme_loader.py

import os
import json
import datasets
from datasets import Dataset, DatasetDict, load_dataset, Features, Value, Image

# Define the paths to your dataset
image_root_dir = "./"
train_jsonl_file_path = "arabic_memes_categorization_train.jsonl"
dev_jsonl_file_path = "arabic_memes_categorization_dev.jsonl"
test_jsonl_file_path = "arabic_memes_categorization_test.jsonl"

# Define features for the dataset
features = Features({
    'id': Value('string'),
    'text': Value('string'),
    'image': Image(),
    'img_path': Value('string')
})

# Function to load each dataset split
def load_armeme_split(jsonl_file_path, image_root_dir):
    data = []
    
    # Load JSONL file
    with open(jsonl_file_path, 'r') as f:
        for line in f:
            item = json.loads(line)
            # Update image path to absolute path
            item['img_path'] = os.path.join(image_root_dir, item['img_path'])
            data.append(item)
    
    # Create a Hugging Face dataset
    dataset = Dataset.from_dict(data, features=features)
    return dataset

# Load each split
train_dataset = load_armeme_split(train_jsonl_file_path, image_root_dir)
dev_dataset = load_armeme_split(dev_jsonl_file_path, image_root_dir)
test_dataset = load_armeme_split(test_jsonl_file_path, image_root_dir)

# Create a DatasetDict
dataset_dict = DatasetDict({
    'train': train_dataset,
    'dev': dev_dataset,
    'test': test_dataset
})

# Push the dataset to Hugging Face Hub
dataset_dict.push_to_hub("QCRI/ArMeme", license="CC-By-NC-SA-4.0")