Column order is nondeterministic when loading from JSON

[Albert Villanova del Moral]

As reported by @meg-huggingface, the order of the JSON object keys is not preserved while loading a dataset from a JSON file with a list of objects.

For example, when loading a JSON files with a list of objects, each with the following ordered keys:

• [ID, Language, Topic],

the resulting dataset may have columns:

• [ID, Topic, Language], or
• [Topic, Language, ID], or
• [Topic, ID, Language],...

This issue is caused by the use of a Python set (which does not preserve the order):

datasets/src/datasets/packaged_modules/json/json.py

Line 168 in 60d21ef

 keys = set().union(*[row.keys() for row in dataset]) 

introduced in

• #5772