A scenario that never reached the model is not scored - #3
Open
veer0608 wants to merge 1 commit into
Open
Conversation
It arrives shaped exactly like one the agent failed: no tools, no answer, scores zero. Nine of eighty scenario-runs died that way on a provider-side 400, and every one of them was averaged into the reported F1, so the number measured the provider rather than the agent. The aggregate now runs over the scenarios that actually reached the model. The rest are counted, named in the summary as NOT SCORED, and marked in the report table rather than quietly printed as failures. A run where nothing reached the model now reports nothing rather than zero. The effect is not small: a run scoring 0.77 under the old accounting scores 0.958 over the eight scenarios that ran.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
A provider outage arrives shaped exactly like an agent failure: no tools, no answer, scores zero. 9 of 80 scenario-runs died that way on a provider-side 400 during yesterday's eight-run series, and every one was averaged into the reported F1. The number measured the provider, not the agent.
Change
scenarios: 10 (8 scored)andNOT SCORED: 2 never reached the model (plan-limits, known-issue-rate-limit).NOT SCOREDinstead of printing them as failures with F1 0.00.run_erroris carried on the row, so the count is recoverable from the JSON scorecard.Effect
A live run that scored 0.77 under the old accounting scores 0.958 over the eight scenarios that ran. The previously published mean of 0.78 across eight runs is therefore a floor, not an estimate.
38 tests, up from 36. The two new ones assert that an errored row does not drag the mean and that the count is reported rather than hidden.