Enhancements for Dataset Handling, Training Splits, Checkpoint Management, and Model Sharing

[MubarakHAlketbi]

• Import Dataset and get_dataset_split_names from datasets to enable more flexible data handling
• Add --train_split and --valid_split arguments in get_args() function for user-defined train and validation splits
• Add --save_limit argument in get_args() function to control the number of model checkpoints saved
• Add local_rank to get the local rank from the environment variable for torchrun, enabling distributed training
• Update prepare_sample_text() function to check if the example is a dictionary, enhancing data compatibility
• Update create_datasets() function to handle train and validation split based on --train_split and --valid_split arguments, providing more flexibility in dataset splitting
• Update run_training() function to include save_strategy, save_total_limit, hub_strategy, and load_best_model_at_end in TrainingArguments, providing more control over model saving and loading
• Add step in run_training() function to push the model to the hub after saving the final checkpoint, facilitating model sharing
• Add print statement to indicate when the model is being pushed to the hub, improving user feedback

[MubarakHAlketbi]

more debugging required