[Feature] Unsloth fine-tuning support

[brragorn]

Feature request

AdOdd4004 and Amazing Q request from Reddit

Motivation / references

toward universal support for fine-tuning

Your contribution

linguistic

OPE-977

[taenin]

Thanks for the feature request! Here's the reddit post referenced in the request for posterity: https://www.reddit.com/r/LocalLLaMA/comments/1id3ak8/comment/m9yq4v4/

[brragorn]

Thanks @taenin
Let's post the update back to that thread if this gets taken up and completed

[georgedouzas]

Hey everyone,

I’d love to help out with this! Just wondering:

• Any key things to keep in mind for integration?
• Is there a preferred approach, or is it open-ended?
• Any blockers or decisions that need sorting first?

[taenin]

@georgedouzas I would recommend starting by understanding how both Oumi and Unsloth ingest datasets and models.

Ultimately to add unsloth support, you'll need to update this method: https://github.com/oumi-ai/oumi/blob/main/src/oumi/builders/training.py#L49

This will involve adding a TrainerType.UNSLOTH. You'll also need to ensure that we can support UNSLOTH with our https://github.com/oumi-ai/oumi/blob/main/src/oumi/core/trainers/base_trainer.py#L21 (or update this class as needed).

It would also be good to understand how our standard training configurations can be ported to inputs for Unsloth: https://oumi.ai/docs/en/latest/api/oumi.core.configs.html#oumi.core.configs.TrainingConfig

[georgedouzas]

Started working on this, please assign it to me. Thanks!