Adding generic preference dataset builder

[SalmanMohammadi]

Context

What is the purpose of this PR? Is it to

• add a new feature
• fix a bug
• update tests and/or documentation
• other (please add here)

I was writing the docs for the preference datasets but noticed we don't have a generic builder to point to, so, here we are.

Changelog

What are the changes made in this PR?

• I've added a preference_dataset builder which uses ChosenToRejectedMessages. I don't think it's worth adding the stack exchange message format as an option at this time.
• I've also exposed the HH RLHF dataset builder in our docs.

Test plan

Please make sure to do each of the following if applicable to your PR. If you're unsure about any one of these just ask and we will happily help. We also have a contributing page for some guidance on contributing.

• run pre-commit hooks and linters (make sure you've first installed via pre-commit install)
• add unit tests for any new functionality
• update docstrings for any new or updated methods or classes
• run unit tests via pytest tests
• run recipe tests via pytest tests -m integration_test
• manually run any new or modified recipes with sufficient proof of correctness
• include relevant commands and any other artifacts in this summary (pastes of loss curves, eval results, etc.)

UX

If your function changed a public API, please add a dummy example of what the user experience will look like when calling it.
Here is a docstring example
and a tutorial example

• I did not change any public API
• I have added an example to docs or docstrings

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchtune/1623

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit adb8934 with merge base c5db813 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[RdoubleA]

do you want to also enable packing for preference datasets? should be easy to test and there's nothing blocking it

[RdoubleA]

nice, we hoenstly need more of these local dataset tests

[RdoubleA]

🫢 🫢 🫢

[RdoubleA]

We should fix this trim whitespace issue at some point...

[SalmanMohammadi]

actually renders for me as 'user\n\nWhat do I do when I have a hole in my trousers?assistant\n\nTake them off.' with llama3 tokenizer and skip_special_tokens=True.

[SalmanMohammadi]

do you want to also enable packing for preference datasets? should be easy to test and there's nothing blocking it

I'm actually pretty out of the loop on packed datasets, sorry. TLDR how to?

[RdoubleA]

I'm actually pretty out of the loop on packed datasets, sorry. TLDR how to?

just emulate the logic here and that's all you need, then expose the packed parameter:

torchtune/torchtune/datasets/_instruct.py

Line 257 in c5db813

if packed:

then just run a DPO config with preference_dataset and use the override dataset.packed=True tokenizer.max_seq_len=4096 and see if it packs successfully and starts training with a few reasonable loss steps. You can hardcode num packs to make it pack faster

You could do this in a follow-up, but would be a nice boost to our RLHF recipes

[SalmanMohammadi]

I'm actually pretty out of the loop on packed datasets, sorry. TLDR how to?

just emulate the logic here and that's all you need, then expose the packed parameter:

torchtune/torchtune/datasets/_instruct.py

Line 257 in c5db813

if packed:

then just run a DPO config with preference_dataset and use the override dataset.packed=True tokenizer.max_seq_len=4096 and see if it packs successfully and starts training with a few reasonable loss steps. You can hardcode num packs to make it pack faster

You could do this in a follow-up, but would be a nice boost to our RLHF recipes

Looks suspiciously straightforward - will follow up with this.