[Feature] Add Tulu3 SFT Mixture Dataset Support

[bwalshe]

Description

This adds support for the Tülu 3 SFT Mixture dataset. (https://huggingface.co/datasets/allenai/tulu-3-sft-mixture)

Currently this PR is a work in progress, as I am unsure what testing process is needed and also what would make a suitable example config for demonstrating the dataset in use. I have included some basic unit tests, and the example config demonstrates the data being used in fine-tuning, but it serves more as a test things are working instead of being a good demonstration of what to do with the data.

Related issues

Fixes #1361

Before submitting

• This PR only changes documentation. (You can ignore the following checks in that case)
• Did you read the contributor guideline Pull Request guidelines?
• Did you link the issue(s) related to this PR in the section above?
• Did you add / update tests where needed?

Reviewers

At least one review from a member of oumi-ai/oumi-staff is required.

[bwalshe]

This config is based on https://github.com/allenai/open-instruct/blob/main/configs/train_configs/sft/mini.yaml

[taenin]

Thanks for the contribution, bwalshe! Overall things look good. I left a few small style comments.

The location you chose for your config is fine for a demonstration. If you've tuned parameters and get good results, I'd add it to the recipes section instead.

[taenin]

One more small issue that came up when running pytest.

[taenin]

Suggested change

	raise ValueError(f"Unknown role {e.args[0]}, was expecting one of {
	roles.keys()}") from e
	raise ValueError(f"Unknown role {e.args[0]}, was expecting one of "
	f"{roles.keys()}") from e

[xrdaukar]

Technically message['role'] can throw another exception if 'role' is not in message.

How about pre-computing the value once at the beginning of the function:
role_str = message["role"].lower().strip()
?