Add "Language Modeling to Unpaired Preference" in Utilities for converting dataset types

[AMindToThink]

This link adds a section to "Dataset Formats" for converting Language Modeling data to Unpaired Preference Data.

My motivation is that I wanted to apply reinforcement learning techniques like BCO to the WMDP dataset: a dataset of for "unlearning" harmful information.

I used the snippet I added to the docs to make an unpaired preference version of the WMDP cyber dataset. Others may want to make similar datasets.

This PR fixes a typo or improves the docs.

[qgallouedec]

Thanks @AMindToThink for the addition.

Strictly speaking, you can't convert a language-modelling dataset (only text column) to a prompt-completion dataset, because you'd have to be able to extract the prompt from it. I'm afraid that adding this part to the documentation will create confusion.

The workaround you use is to have an empty prompt column. Which is a bit strange.

How about instead making the algorithms natively support this new type in which you only have a text column and a label colone?