Skip to content

Group GRPO reward normalization by prompt instead of reshaping the batch - #2

Open
FunJim wants to merge 2 commits into
mainfrom
fix/grpo-reward-normalization-group-collapse
Open

Group GRPO reward normalization by prompt instead of reshaping the batch#2
FunJim wants to merge 2 commits into
mainfrom
fix/grpo-reward-normalization-group-collapse

Reduce the GRPO baseline to one entry per attempt, not per sample

11e9804
Select commit
Loading
Failed to load commit list.

Select a check to view from the sidebar