Shuffle pre-training data

[tmylla]

Running the script before executing the 'Prepare the dataset' step in the README instructions can avoid unexpected fluctuations in the loss curve. This ensures the data is properly shuffled, avoiding the instability observed in the loss curve.

python -m prepare_dataset.shuffle --directory_path your_input_directory_path--output_directory your_output_directory_path--batch_size 10000