Repository navigation
fix(examples): complete Arena metrics and generation defaults - #1746
Merged
Merged
Conversation
Record generated-turn distributions and bounded harness outcome rates for exported episodes so rollout quality changes can be diagnosed. Keep reward assignment and failure attribution unchanged.
Link exported interactions to Arena tasks and proxy sessions. Report attempted-group quality by task type and stream while retaining existing metric names. Tolerate missing measurements and invalid optional metadata without fabricating rewards or changing failure attribution.
Le8r0nJames
requested review from
CormickKneey,
HwVanICI,
PrometheusComing,
fishcrap,
garrett4wade and
sitabulaixizawaluduo
as code owners
September 22, 2026 06:50
This was referenced Sep 22, 2026
4 of 15 tasks
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Description
Complete Arena rollout observability and forward configured sampling defaults into proxy sessions.
top_p/top_kalongside temperature and the token limit. Explicit request fields and token-limit aliases retain precedence.Harness rates cover usable exported episodes; group ratios cover attempted groups. Turn counts come from the last exported interaction (cumulative for concat exports). This adds no subprocess-agent metadata transport. Reward assignment, failure classification and filtering policies are unchanged.
Related Issue
Follow-up to #1702; the outcome metrics also extend #1724, and default precedence is related to #1725. No issue is closed by this PR.
Type of Change
Checklist
pre-commit run --all-files)main/review-prcommandAdditional Context
Combined CPU regression: 205 passed against main
c74b343c. Configuration consistency and CLI generation also passed in the target image. Read-only integration review found no blocking issues.Full host pre-commit was run: formatting, lint, license and other static hooks passed; the configuration-consistency and CLI-generation hooks are blocked by host Python 3.10 / older PyTorch. Both scripts pass in the target Python 3.12 / PyTorch 2.9.1 image. No GPU end-to-end validation is claimed for this branch.