Fix KL approximation and evaluation prompt-response alignment#7
Open
Tianqi-Xuuu wants to merge 1 commit intollmsystem:mainfrom
Open
Fix KL approximation and evaluation prompt-response alignment#7Tianqi-Xuuu wants to merge 1 commit intollmsystem:mainfrom
Tianqi-Xuuu wants to merge 1 commit intollmsystem:mainfrom
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR fixes the KL approximation direction and corrects prompt-response alignment in evaluation so all sampled responses are scored.
For KL, the previous implementation used the approximation in the wrong direction, so the computed KL did not correctly match the intended
KL(policy || ref)quantity. This change makes the KL computation consistent with the current policy being compared against the reference policy.For evaluation,
evaluate_policygenerates multiple responses per prompt, but the old code only paired responses with the original prompt list. As a result, only part of the generated responses were actually scored. This change uses the duplicated prompt list so every sampled response is included in reward evaluation.