Gemma3 SFT fails to preprocess a custom JSONL dataset (NoneType not iterable) and ends up with an empty train set

Env:

Training setup: 8 GPUs
ms-swift: 3.9.1
deepspeed: 0.17.4
transformers:4.55.0

Description:
I’m fine-tuning Gemma3 Instruct using swift sft with a custom JSONL dataset.
With the same training command and the same dataset, switching only --model to Qwen3 works, but Gemma3 fails during dataset preprocessing / template encoding.

Error:
[WARNING:swift] 👆👆👆There are errors in the dataset, the data will be deleted
[INFO:swift] Traceback (most recent call last):
File "/home/tiger/.local/lib/python3.11/site-packages/swift/llm/dataset/preprocessor/core.py", line 178, in batched_preprocess
row = self.preprocess(row)
^^^^^^^^^^^^^^^^^^^^
File "/home/tiger/.local/lib/python3.11/site-packages/swift/llm/dataset/utils.py", line 289, in preprocess
encoded = self.template.encode(row, return_length=True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/tiger/.local/lib/python3.11/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/home/tiger/.local/lib/python3.11/site-packages/swift/llm/template/base.py", line 497, in encode
encoded = self._encode_truncated(chosen)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/tiger/.local/lib/python3.11/site-packages/swift/llm/template/base.py", line 1201, in _encode_truncated
encoded = self._encode(inputs)
^^^^^^^^^^^^^^^^^^^^
File "/home/tiger/.local/lib/python3.11/site-packages/swift/llm/template/base.py", line 1229, in _encode
self._swift_encode(inputs) if template_backend == 'swift' else self._jinja_encode(inputs))
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/tiger/.local/lib/python3.11/site-packages/swift/llm/template/template/gemma.py", line 90, in _swift_encode
return super()._swift_encode(inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/tiger/.local/lib/python3.11/site-packages/swift/llm/template/base.py", line 1085, in _swift_encode
self._concat_context_list(prefix, res_context_list, res_context_types, system=system)
File "/home/tiger/.local/lib/python3.11/site-packages/swift/llm/template/base.py", line 690, in _concat_context_list
for context in context_list:
TypeError: 'NoneType' object is not iterable

[rank0]: Traceback (most recent call last):
[rank0]: File "/home/tiger/.local/lib/python3.11/site-packages/swift/cli/sft.py", line 10, in
[rank0]: sft_main()
[rank0]: File "/home/tiger/.local/lib/python3.11/site-packages/swift/llm/train/sft.py", line 340, in sft_main
[rank0]: return SwiftSft(args).main()
[rank0]: ^^^^^^^^^^^^^^^^^^^^^
[rank0]: File "/home/tiger/.local/lib/python3.11/site-packages/swift/llm/base.py", line 49, in main
[rank0]: result = self.run()
[rank0]: ^^^^^^^^^^
[rank0]: File "/home/tiger/.local/lib/python3.11/site-packages/swift/llm/train/sft.py", line 169, in run
[rank0]: train_dataset, val_dataset = self._prepare_dataset()
[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^
[rank0]: File "/home/tiger/.local/lib/python3.11/site-packages/swift/llm/train/sft.py", line 131, in _prepare_dataset
[rank0]: datasets = self._post_process_datasets(datasets)
[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]: File "/home/tiger/.local/lib/python3.11/site-packages/swift/llm/train/sft.py", line 164, in _post_process_datasets
[rank0]: self._show_dataset(*datasets)
[rank0]: File "/home/tiger/.local/lib/python3.11/site-packages/swift/llm/train/sft.py", line 292, in _show_dataset
[rank0]: inputs = train_dataset[0] if hasattr(train_dataset, 'len') else next(iter(train_dataset))
[rank0]: ~~~~~~~~~~~~~^^^
[rank0]: File "/home/tiger/.local/lib/python3.11/site-packages/datasets/arrow_dataset.py", line 2780, in getitem
[rank0]: return self._getitem(key)
[rank0]: ^^^^^^^^^^^^^^^^^^
[rank0]: File "/home/tiger/.local/lib/python3.11/site-packages/datasets/arrow_dataset.py", line 2764, in _getitem
[rank0]: pa_subtable = query_table(self._data, key, indices=self._indices)
[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]: File "/home/tiger/.local/lib/python3.11/site-packages/datasets/formatting/formatting.py", line 593, in query_table
[rank0]: _check_valid_index_key(key, size)
[rank0]: File "/home/tiger/.local/lib/python3.11/site-packages/datasets/formatting/formatting.py", line 533, in _check_valid_index_key
[rank0]: raise IndexError(f"Invalid key: {key} is out of bounds for size {size}")
[rank0]: IndexError: Invalid key: 0 is out of bounds for size 0

Reproduce the Gemma3 SFT run with the custom JSONL dataset, then inspect swift/llm/template/template/gemma.py and the encoding flow in swift/llm/template/base.py. Follow how preprocessing in swift/llm/dataset/preprocessor/core.py handles the failed row and how swift/llm/train/sft.py reports the dataset. Done means valid input no longer raises the NoneType error and the resulting train set is nonempty.