Skip to content

Align validation step with checkpoint step#54

Open
licesma wants to merge 1 commit into
physical-superintelligence-lab:mainfrom
licesma:fix/align_validation_with_checkpoint
Open

Align validation step with checkpoint step#54
licesma wants to merge 1 commit into
physical-superintelligence-lab:mainfrom
licesma:fix/align_validation_with_checkpoint

Conversation

@licesma

@licesma licesma commented Jun 3, 2026

Copy link
Copy Markdown
Contributor

Validation was triggered on global_step % validation_steps == 0, while checkpointing
uses (global_step + 1) % checkpointing_steps == 0. This one-step offset means the
eval loss logged near a checkpoint does not correspond to that checkpoint's weights.

As a side effect, validation no longer runs at step 0 (on the barely-trained model right
after the first step)

@songlin songlin left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It is helpful to run the validation code path for the first training loop. Can you revise the code again?

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