Support for images_root field for multimodal training data

[tianbinli]

Reminder

• I have read the above rules and searched the existing issues.

Description

Hi, thanks for the great repo!

Could you consider adding support for an images_root field in multimodal training? This would allow specifying a common root directory for image paths, so we can use relative paths in the dataset.

It would simplify data management and improve portability.

Thanks!

Pull Request

No response

[hiyouga]

Please use

LLaMA-Factory/src/llamafactory/hparams/data_args.py

Lines 42 to 45 in 2b7d564

	media_dir: Optional[str] = field(
	default=None,
	metadata={"help": "Path to the folder containing the images, videos or audios. Defaults to `dataset_dir`."},
	)

[tianbinli]

Thank you for the quick response! I’ve figured it out — it's possible to set a media_dir for overall datasets, which is really helpful. However, if there are many datasets with different root paths, it seems like this setup doesn’t fully support that scenario yet.

Please use

LLaMA-Factory/src/llamafactory/hparams/data_args.py

Lines 42 to 45 in 2b7d564

media_dir: Optional[str] = field(
default=None,
metadata={"help": "Path to the folder containing the images, videos or audios. Defaults to dataset_dir."},
)

[hiyouga]

make sense, we'll add some features in the future