🍟 [SFT] Handles the dataset if it has been preprocessed

[BenasdTW]

What does this PR do?

Fixes #2861

Adding back the code in SFTTrainer that properly handles the dataset if it has been preprocessed.
This part of the code was removed in #2405, but it should not have been removed.

        # If the dataset is already preprocessed (tokenized), return as-is. Only works if dataset is
        # a datasets.Dataset or datasets.IterableDataset -- not for torch Dataset
        column_names = (
            dataset.column_names if isinstance(dataset, (Dataset, IterableDataset)) else None
        )
        if column_names and "input_ids" in column_names:
            if formatting_func is not None:
                warnings.warn(
                    "You passed a dataset that is already processed (contains an `input_ids` field) together with a "
                    "valid formatting function. Therefore `formatting_func` will be ignored. Either remove the "
                    "`formatting_func` or pass a dataset that is not already processed.",
                    UserWarning,
                )
            if not packing:
                return dataset

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.

[qgallouedec]

Thanks for fixing it!
I made a few comments. Can you also add a test?

[BenasdTW]

Added test: test_sft_trainer_directly_with_pretokenized_data

root@813ecfcb235b:/workspaces/LLMTrain/trl# python -m pytest tests/test_sft_trainer.py::SFTTrainerTester::test_sft_trainer_directly_with_pretokenized_data
=================================================== test session starts ====================================================
platform linux -- Python 3.11.10, pytest-8.3.4, pluggy-1.5.0
rootdir: /workspaces/LLMTrain/trl
configfile: pyproject.toml
plugins: hypothesis-6.115.5, rerunfailures-15.0, cov-6.0.0, xdist-3.6.1, anyio-4.8.0
collected 1 item                                                                                                           

tests/test_sft_trainer.py .                                                                                          [100%]

==================================================== 1 passed in 18.36s ====================================================

[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.

[qgallouedec]

Thanks @BenasdTW ! can be merged as soon as the CI is green

[hanyin88]

Thanks for the kind contribution. Wondering why

def tokenize(ex):
                    tokenized = processing_class(ex[args.dataset_text_field])
                    return {"input_ids": tokenized["input_ids"], "attention_mask": tokenized["attention_mask"]}

dataset = dataset.map(tokenize, **map_kwargs)

would generate new dataset fingerprint now, and not be able to utilize prior cache? All prior steps could utilize the cache with same fingerprint without problem. Many thanks.