Well to be honest, we kind of thought it's redundant.
Basically storing artifacts in experiments and having them retrieved quickly from the code itself was way more convenient for us then to manually have to do clone/pull of the data...
Example:
Create dataset:

task = Task.init("Dataset Sample", "version 1") task.upload_artifact("dataset_folder", './my_local_folder') print('task.id = {}'.format(task.id))
Get dataset in runtime:

dataset_task = Task.get_task(task_id='my_task_id') my_local_path = dataset_task.artifacts['dataset_folder'].get()