Skip to content

Conversation

@TomeHirata
Copy link
Collaborator

@TomeHirata TomeHirata commented Nov 12, 2025

Instead of relying on "capture_trace" option, enable/disable logging based on the sample size since sample size is a more accurate indicator.

@LakshyAAAgrawal
Copy link
Collaborator

Hi @TomeHirata, one thing to be aware of is that GEPA will soon have support for not running eval on len(valset) but instead a subset of it, similar to what MIPRO does, hence this scheme may stop working then. I am trying to think what might be a good solution, maybe tracking reflection_minibatch_size may be better here?

@TomeHirata
Copy link
Collaborator Author

@LakshyAAAgrawal thanks for sharing the plan, so should we disable logging when len(valset)== reflection_minibatch_size?

@LakshyAAAgrawal
Copy link
Collaborator

I think len(valset) <= reflection_minibatch_size should be good.

Updated the DspyAdapter and GEPA classes to replace the full_eval_size parameter with reflection_minibatch_size for improved clarity. Adjusted the evaluate method and corresponding tests to reflect this change, ensuring callback metadata is correctly generated based on the new parameter.

Signed-off-by: TomuHirata <[email protected]>
@TomeHirata
Copy link
Collaborator Author

@LakshyAAAgrawal sounds good, updated

@TomeHirata TomeHirata merged commit 4dd085c into stanfordnlp:main Nov 13, 2025
12 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants