GSPO vs GRPO/DAPO #3335
Unanswered
rhinosaur0
asked this question in
Q&A
GSPO vs GRPO/DAPO
#3335
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
GSPO theoretically makes much more sense than GRPO/DAPO, weighing each token in the sequence equally rather than weighing them differently due to individual log-prob. If so, how was the "noisy" GRPO still able to achieve results, while GSPO didn't achieve that much better results?
I'm curious to hear some explanations!
Beta Was this translation helpful? Give feedback.
All reactions