fix: need to pass skip_prepare_dataset for pretokenized dataset due to breaking change in HF SFTTrainer

[HarikrishnanBalagopal]

Description of the change

Fix test failure

Related issue number

Fixes #324 (comment)

How to verify the PR

make test

Was the PR tested

• I have added >=1 unit test(s) for every new method I have added.
• I have ensured all unit tests pass

[willmj]

This seems to be the line in the SFTTrainer code that caused the error on our side

[willmj]

LGTM! I'll try to get another set of eyes on this before merging though.

[anhuong]

The issue is that trl created a release 5 days ago that our PRs are now picking up. v0.10.1 includes this line change https://github.com/huggingface/trl/blob/d57e4b726561e5ae58fdc335f34029052944a4a3/trl/trainer/sft_trainer.py#L348 where if packing=False (set by default in our code) then it expects that data preparation must be completed. For pretokenized datasets, data preparation is not needed, thus we need to set skip_prepare_dataset=False. For the other dataset formats we accept, dataset_text_field is set and thus data processing is run in SFTTrainer.

Angel had the idea that we could only add skip_prepare_dataset=False in our pretokenized dataset tests only. I don't think we could as this is not something the user can set, this is something we need to pass into SFTTrainer for all pretokenized datasets.