fix(eval): deduplicate queries and qualify explicit evidence - #655
Merged
Conversation
Group relevance sets by query and label source, exclude every citing anchor, reject unknown source labels, and compare retrieval strategies only on explicit references. Generated with [p/](https://www.prjct.app/)
… explicit evidence (round 2 of 3) (v4.22.0) Generated with [p/](https://www.prjct.app/)
jlopezlira
changed the title
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Retrieval evaluation counted one query with several labeled results as several independent samples and repeated the same served lookup. It now groups positives into one query case, excludes every citing anchor, and keeps explicit references separate from automatic proxy labels.
Comparison gates use distinct explicit-reference queries only; unknown label sources cannot silently become explicit evidence. Reports describe date cohorts over the current corpus and no longer imply a historical replay.
Validation: focused evaluation and integration tests pass (52 tests, 115 assertions). The reproduction changes three identical-query lookups into one while retaining all relevant results. On this repository the served evaluation drops from 131 lookups to 14 distinct queries; the explicit-reference metrics remain unchanged. This is reduced evaluation work, not a claim of provider-token savings or improved model outcomes.
Round 2 of the requested three-round ship/merge/upgrade/evaluation cycle.