Fix StopIteration when the JSON data files are empty

[LeSingh1]

Loading a JSON dataset whose data file is empty fails with a bare StopIteration:

open("empty.jsonl", "w").close()
load_dataset("json", data_files="empty.jsonl", split="train", streaming=True)
# StopIteration

File "src/datasets/packaged_modules/json/json.py", line 101, in _split_generators
StopIteration

_split_generators() infers the features from the first table of the first split:

if self.info.features is None:
    try:
        pa_table = next(iter(self._generate_tables(**splits[0].gen_kwargs, allow_full_read=False)))[1]
        ...
    except FullReadDisallowed:
        pass

Only FullReadDisallowed is caught. When the data files hold no data the generator yields nothing, so next() raises StopIteration and it propagates out of load_dataset — an exception with no message that says nothing about the file being empty. It is also the kind of exception that gets swallowed or converted to RuntimeError (PEP 479) depending on where the call ends up.

The other loaders don't do this — text streams an empty file as zero examples.

Fix

Pass a default to next() and only set the features when a table was actually produced. Behaviour afterwards:

	before	after
streaming=True	StopIteration	yields no example, like the text loader
streaming=False	StopIteration	the usual "specify features" DatasetGenerationError, like csv / text

A file containing only newlines is unaffected — it already produces a zero-row table and infers features={}, so it never hit this path.

Tests

• test_json_split_generators_with_empty_file — _split_generators() leaves info.features as None and generates no table.
• test_json_load_dataset_streaming_empty_file — end to end, streaming an empty file yields [].

Both fail on main with StopIteration.