[Feature] Exploring support for TextGrad and DSPy libraries

[georgedouzas]

Feature request

Has there been any thought about supporting libraries like TextGrad or DSPy?

Motivation / references

Both projects focus on making prompt-based workflows more structured and optimized, and it seems like they could fit really well with Oumi, especially for more advanced training, fine-tuning, or reasoning-based tasks.

Your contribution

Happy to explore or help implement if there’s interest.

[kaisopos]

Oumi does not have support for prompt engineering and prompt optimization at the moment, so both additions seem to be very useful for the platform. That said, I am not sure how much work is involved to integrate with these libraries, it seems to be a fair amount of effort. Is this something that you would be willing to scope and implement yourself?

[georgedouzas]

I think TextGrad focuses on two things: text instance optimization and prompt optimization. These directly impact Oumi’s inference, evaluation, and LLM judge features. I can start by wrapping the main TextGrad functionalities and providing them as independent features, then figure out how to connect them to Oumi’s main API. What do you think?

[kaisopos]

@oelachqar any thoughts?

[kaisopos]

Quickly synced with a few folks at Oumi. We all agree that this is useful, but might be complex.
Do you want to start with a notebook for a proof-of-concept, before you send any PRs that need proper code reviewing, tests, etc?