Default dataset_text_field to "text"

[qgallouedec]

What does this PR do?

Most datasets use “text” for the text field. Moreover, this is the column we document in trl as [standard language modeling] (https://huggingface.co/docs/trl/en/dataset_formats#overview-of-the-dataset-formats-and-types).
Consequently, it makes sense to use this column by default for dataset_text_field.
On the other hand, it allows us to drop # needs dummy text field as here:

trl/examples/scripts/sft_vlm.py

Line 57 in 78249d9

training_args.dataset_text_field = "" # need a dummy field

Implications

• The dataset_text_field can no longer be None in the SFTConfig. We then use the following logic in SFTTrainer: if formatting_func is provided, use it, otherwise, use dataset_text_field.
• Things are a little different in ConstantLengthDataset where formatting_func and dataset_text_field are None by default, and only one of them can be supplied.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a GitHub issue? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[lewtun]

not super important, but maybe we should type the rest of the function signature for consistency?

[qgallouedec]

Agree in general, but I didn't want to spend to much time on it since I plan to refactor SFT data preprocessing, to be more aligned with what's done in other trainers (which is clearer IMO)