sft微调qwen3.5 构造自己的VQA数据集jsonl ，报错An error occurred while generating the dataset

Checklist / 检查清单

• I have searched existing issues, and this is a new bug report. / 我已经搜索过现有的 issues，确认这是一个新的 bug report。

Bug Description / Bug 描述

[INFO:swift] default_system: None
[INFO:swift] max_length: 2048
[INFO:swift] response_prefix: '<think>\n'
[INFO:swift] agent_template: qwen3_5
[INFO:swift] norm_bbox: norm1000
[INFO:swift] Setting ROOT_IMAGE_DIR: None. You can adjust this hyperparameter through the environment variable: `ROOT_IMAGE_DIR`.
[INFO:swift] Setting QWENVL_BBOX_FORMAT: legacy. You can adjust this hyperparameter through the environment variable: `QWENVL_BBOX_FORMAT`.
[INFO:swift] Start time of running main: 2026-03-13 17:15:58.510713
[INFO:swift] swift.__version__: 4.1.0.dev0
[INFO:swift] SelfCognitionPreprocessor has been successfully configured with name: ('swift-robot', 'swift-robot'), author: ('swift', 'swift').
Setting num_proc from 4 back to 1 for the train split to disable multiprocessing as it only contains one shard.

Generating train split: 0 examples [00:00, ? examples/s]Failed to load JSON from file '/appdata/zhangkailin/Qwen2.5-VL/data_process_genn/process/result/safety_vqa_data_v6_1223_50.json_swift0313.jsonl' with error <class 'pyarrow.lib.ArrowInvalid'>: JSON parse error: Column(/messages/[]/content) changed from string to object in row 0

Generating train split: 0 examples [00:00, ? examples/s]
[rank0]: Traceback (most recent call last):
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/datasets/packaged_modules/json/json.py", line 160, in _generate_tables
[rank0]:     df = pandas_read_json(f)
[rank0]:          ^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/datasets/packaged_modules/json/json.py", line 38, in pandas_read_json
[rank0]:     return pd.read_json(path_or_buf, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/pandas/io/json/_json.py", line 815, in read_json
[rank0]:     return json_reader.read()
[rank0]:            ^^^^^^^^^^^^^^^^^^
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/pandas/io/json/_json.py", line 1014, in read
[rank0]:     obj = self._get_object_parser(self.data)
[rank0]:           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/pandas/io/json/_json.py", line 1040, in _get_object_parser
[rank0]:     obj = FrameParser(json, **kwargs).parse()
[rank0]:           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/pandas/io/json/_json.py", line 1176, in parse
[rank0]:     self._parse()
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/pandas/io/json/_json.py", line 1392, in _parse
[rank0]:     ujson_loads(json, precise_float=self.precise_float), dtype=None
[rank0]:     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]: ValueError: Trailing data

[rank0]: During handling of the above exception, another exception occurred:

[rank0]: Traceback (most recent call last):
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/datasets/builder.py", line 1855, in _prepare_split_single
[rank0]:     for _, table in generator:
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/datasets/packaged_modules/json/json.py", line 163, in _generate_tables
[rank0]:     raise e
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/datasets/packaged_modules/json/json.py", line 137, in _generate_tables
[rank0]:     pa_table = paj.read_json(
[rank0]:                ^^^^^^^^^^^^^^
[rank0]:   File "pyarrow/_json.pyx", line 342, in pyarrow._json.read_json
[rank0]:   File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
[rank0]:   File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
[rank0]: pyarrow.lib.ArrowInvalid: JSON parse error: Column(/messages/[]/content) changed from string to object in row 0

[rank0]: The above exception was the direct cause of the following exception:

[rank0]: Traceback (most recent call last):
[rank0]:   File "/appdata/zhangkailin/ms-swift-new/swift/cli/sft.py", line 20, in <module>
[rank0]:     sft_main()
[rank0]:   File "/appdata/zhangkailin/ms-swift-new/swift/pipelines/train/sft.py", line 354, in sft_main
[rank0]:     return SwiftSft(args).main()
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/appdata/zhangkailin/ms-swift-new/swift/pipelines/base.py", line 52, in main
[rank0]:     result = self.run()
[rank0]:              ^^^^^^^^^^
[rank0]:   File "/appdata/zhangkailin/ms-swift-new/swift/ray/base.py", line 168, in wrapper
[rank0]:     return func(self, *args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/appdata/zhangkailin/ms-swift-new/swift/pipelines/train/sft.py", line 174, in run
[rank0]:     train_dataset, val_dataset = self._prepare_dataset()
[rank0]:                                  ^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/appdata/zhangkailin/ms-swift-new/swift/ray/base.py", line 168, in wrapper
[rank0]:     return func(self, *args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/appdata/zhangkailin/ms-swift-new/swift/pipelines/train/sft.py", line 120, in _prepare_dataset
[rank0]:     train_dataset, val_dataset = self._get_dataset()
[rank0]:                                  ^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/appdata/zhangkailin/ms-swift-new/swift/pipelines/train/sft.py", line 84, in _get_dataset
[rank0]:     train_dataset, val_dataset = load_dataset(
[rank0]:                                  ^^^^^^^^^^^^^
[rank0]:   File "/appdata/zhangkailin/ms-swift-new/swift/dataset/loader.py", line 327, in load_dataset
[rank0]:     train_dataset = loader.load(dataset_syntax, dataset_meta, use_hf=use_hf)
[rank0]:                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/appdata/zhangkailin/ms-swift-new/swift/dataset/loader.py", line 159, in load
[rank0]:     dataset = self._load_dataset_path(
[rank0]:               ^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/appdata/zhangkailin/ms-swift-new/swift/dataset/loader.py", line 55, in _load_dataset_path
[rank0]:     dataset = hf_load_dataset(file_type, data_files=dataset_path, **kwargs)
[rank0]:               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/datasets/load.py", line 2084, in load_dataset
[rank0]:     builder_instance.download_and_prepare(
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/datasets/builder.py", line 925, in download_and_prepare
[rank0]:     self._download_and_prepare(
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/datasets/builder.py", line 1001, in _download_and_prepare
[rank0]:     self._prepare_split(split_generator, **prepare_split_kwargs)
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/datasets/builder.py", line 1742, in _prepare_split
[rank0]:     for job_id, done, content in self._prepare_split_single(
[rank0]:   File "/home/zhangkailin/.conda/envs/swift3/lib/python3.12/site-packages/datasets/builder.py", line 1898, in _prepare_split_single
[rank0]:     raise DatasetGenerationError("An error occurred while generating the dataset") from e
[rank0]: datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset
[rank0]:[W313 17:16:00.670177029 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

How to Reproduce / 如何复现

latex-ocr 开源数据集没问题，自己构造的vqa报错，参考https://swift.readthedocs.io/zh-cn/latest/Customization/Custom-dataset.html

Additional Information / 补充信息

No response

Start with the custom JSONL file and compare its messages/content structure with the Custom dataset documentation. Then trace dataset loading through swift/dataset/loader.py and the sft.py entry point, using the reported ArrowInvalid error as the first check. Done means the supplied VQA JSONL loads as a train split without DatasetGenerationError.