Allow pre-tokenized datasets in SFTTrainer

[BramVanroy]

This tiny PR makes it so that datasets that are already tokenized (have an input_ids column) can be used as-is. This greatly improves flexibility, especially in larger training runs: we can do data preprocessing (tokenization) one one CPU-heavy server and save the tokenized dataset to disk, and transfer it to a GPU server to use it directly as a dataset.

Note: this only works for datasets datasets, not for PyTorch datasets.

[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.

[younesbelkada]

Thanks for adding the pre-tokenized dataset support in TRL SFTTrainer ! Can you also add few lines in the docs in a follow up PR ? 🙏 Otherwise happy to do it !

[BramVanroy]

@younesbelkada Sure! #1521

[younesbelkada]

Thanks @BramVanroy !