[Feature] Package to inject Spurious Correlation(s) in huggingface datasets

[MarcelMatsal]

Description

This code introduces functionality for injecting spurious correlations (SSTI from our paper) into huggingface datasets. This will allow us to see how these correlations affect the pretraining of models. It will be expanded to other modalities. Currently only textual injections are possible but soon will add functionality to add injections into image data.

Added an additional dependency of "termcolor" to the dataset additional dependencies

Checklist

• I have read the Contributing document.
• The documentation is up-to-date with the changes I made (check build artifacts).
• All tests passed, and additional code has been covered with new tests.
• I have added the PR to the RELEASES.rst file.

[huguesva]

thanks a lot @MarcelMatsal , are there some plans to include examples of pretraining or finetuning of LLMs later in the library? @RandallBalestriero don't hesitate to review as well if you have time since you know more than me

[MarcelMatsal]

@huguesva Our current plan was to finetune/pretrain some VLMs like CLIP with spurious data and I could include those examples in this library. I could definitely also include some finetuning examples of pure LLMs down the line or we could fully do the pretaining step

[huguesva]

Thanks @MarcelMatsal ! @RandallBalestriero do you validate this PR ? thanks