Replace Claude Safety Judge with Llama3.1 #2974
Merged
Conversation
This file contains 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
farzaank
commented
Sep 4, 2024
| @@ -2,7 +2,7 @@ | |||
|
|
|||
| from helm.benchmark.adaptation.request_state import RequestState | |||
| from helm.benchmark.annotation.annotator import Annotator | |||
| from helm.benchmark.annotation.score_util import score_with_reasoning | |||
| from helm.benchmark.annotation.model_as_judge import score_with_reasoning | |||
There was a problem hiding this comment.
missed this in the last PR for some files, so here it is
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
Replaces Claude with Llama 3.1.
Motivation is that Claude would refuse judging for some model-scenario pairs. We decided to try 2 open models as a judge as an alternative: Llama Guard 3 and Llama 3.1 if that didn't work out.
Llama Guard 3, has the behavior of occasionally rating refusals as unsafe when they mention unsafe subject matter (this false positive rate is well documented: "while deploying Llama Guard 3 will likely improve the safety of your system, it might increase refusals to benign prompts (False Positives)") so to keep behavioral consistency with our other judge (GPT-4o), we use Llama 3.1 instead. For now, I set the size to the largest one.