Training the Qwen3.5-9B model with only_text data, how to avoid multimodal placeholders

[Ye-XX]

When fine-tuning the Qwen3.5-9B model, if the "input" or "output" fields in the data contain multimedia placeholders such as “", an error will occur.

[rank0]: Traceback (most recent call last):
[rank0]:   File "/app/src/llamafactory/launcher.py", line 185, in <module>
[rank0]:     run_exp()
[rank0]:   File "/app/src/llamafactory/train/tuner.py", line 173, in run_exp
[rank0]:     _training_function(config={"args": args, "callbacks": callbacks})
[rank0]:   File "/app/src/llamafactory/train/tuner.py", line 141, in _training_function
[rank0]:     run_sft(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
[rank0]:   File "/app/src/llamafactory/train/sft/workflow.py", line 52, in run_sft
[rank0]:     dataset_module = get_dataset(template, model_args, data_args, training_args, stage="sft", **tokenizer_module)
[rank0]:                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/app/src/llamafactory/data/loader.py", line 318, in get_dataset
[rank0]:     train_dict["train"] = _get_preprocessed_dataset(
[rank0]:                           ^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/app/src/llamafactory/data/loader.py", line 255, in _get_preprocessed_dataset
[rank0]:     dataset = dataset.map(
[rank0]:               ^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.11/site-packages/datasets/arrow_dataset.py", line 560, in wrapper
[rank0]:     out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
[rank0]:                                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.11/site-packages/datasets/arrow_dataset.py", line 3309, in map
[rank0]:     for rank, done, content in iflatmap_unordered(
[rank0]:   File "/opt/conda/lib/python3.11/site-packages/datasets/utils/py_utils.py", line 626, in iflatmap_unordered
[rank0]:     [async_result.get(timeout=0.05) for async_result in async_results]
[rank0]:   File "/opt/conda/lib/python3.11/site-packages/datasets/utils/py_utils.py", line 626, in <listcomp>
[rank0]:     [async_result.get(timeout=0.05) for async_result in async_results]
[rank0]:      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.11/site-packages/multiprocess/pool.py", line 774, in get
[rank0]:     raise self._value
[rank0]: ValueError: The number of images does not match the number of <image> tokens in 
[{'content': "......<image>......", 'role': 'user'}, {'content': ''.....<image>......", 'role': 'assistant'}].

I just want to conduct only_text training.

How can I avoid this issue?

Also, does this problem exist for inference as well?

[identity-wael]

Here <image> is being interpreted by LlamaFactory's multimodal preprocessor as a media placeholder, even though you intended it as literal text. This error occurs before the training step; “text-only dataset” does not automatically escape that marker.

There is a configurable placeholder in the current source: IMAGE_PLACEHOLDER is read from the environment, defaulting to <image>. If your installed version has that setting, you can reserve a different marker that never appears in your text:

IMAGE_PLACEHOLDER='__LF_RESERVED_IMAGE_SLOT_9C6E__'   llamafactory-cli train your_config.yaml

Set it before starting/importing LlamaFactory, and ensure it reaches all training processes. Rebuild any previously tokenized/preprocessed dataset cache. This changes LlamaFactory's input placeholder convention; it does not remove the model's vision components or change its native special tokens. Keep your model-appropriate template.

An alternative is to transform literal <image> strings in the dataset to an unambiguous textual representation, if changing that content is acceptable. Simply wrapping <image> in Markdown backticks won't help a raw substring check.

For inference: the LlamaFactory HF chat engine also calls template.mm_plugin.process_messages, so use the same placeholder convention there. Other serving stacks have their own preprocessing rules; this is not necessarily a model-wide restriction. Which LlamaFactory version and template are you using?