Repository navigation
Retry transient errors in the ta serve eval client - #132
Merged
Merged
Conversation
EndpointEvaluator had no retry, and the base Evaluator turns any exception into an empty prediction. A RunPod proxy outage during the published eval run blanked 35 consecutive LibriSpeech test-other rows, scoring them 100% WER and reporting 6.38 instead of ~3.1. Retry 429/5xx (incl. the proxy's 524 timeout) and httpx transport errors with exponential backoff (8 attempts, ~90 s), mirroring SmallestEvaluator. Client errors like 401 still fail immediately. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
4 of 5 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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Why
ta evalagainst ata serveendpoint had no retry, and the baseEvaluatorscores any exception as an empty prediction. In the published run (20261009_070543_…_librispeech-other), a RunPod proxy outage blanked 35 rows in two contiguous bursts (#405–437, #556–559). Every one was transcribed perfectly in the earlier run on the same pod. Those rows scored 100% WER and took test-other from ~3.1 to 6.38. That 6.38 is the number in the README/MODEL_CARD from e9523ff.No other dataset in that run was affected: the empty predictions on AMI, GigaSpeech and Earnings22 all match empty or filler references.
What
EndpointEvaluator._postretries 429/500/502/503/504/524 andhttpx.TransportErrorwith exponential backoff (8 attempts, ~90 s), the same approachSmallestEvaluatoralready uses.Tests
Follow-up
Re-run
librispeech-otheragainst the pod and correct the published test-other WER, the 12-set mean (~8.44, down from 8.71) and the pooled number.🤖 Generated with Claude Code