Repository navigation
fix(eval): score what a reply means, not how it is typed - #57
Merged
Merged
Conversation
The gpt-6-luna vs gpt-4.1-mini comparison scored Luna 19-21/26 and mini
23-25/26; read by hand, every reply was right. The scorer was wrong:
- Replies are normalised before matching: curly quotes straightened and "n't"
spelt " not", so "isn’t listed" and "aren’t on file" meet the existing
"not listed" / "not on file" patterns. "does not have any mention" added.
- "How long has he been at iEDI?" takes a start month (`since`) instead of a
fixed "11 months": the accepted durations are worked out on the day
(duration_phrases), and --rescore judges each saved run as of the date in
its filename, so August's correct "11 months" stays correct.
- Q19 accepts "has not had mentoring"; Q24 and Q26 accept the substance in
other words ("regular", "daily"; "assembl…", "left the warehouse").
Re-scored, both models get 26/26 on every pass; no reply that was "ok" before
is marked wrong now, in any of the 31 saved runs. New tests:
tests/test_eval_scoring.py.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The gpt-6-luna vs gpt-4.1-mini comparison scored Luna 19–21/26 and mini 23–25/26; read by hand, every reply was correct. Scorer fixes:
n't→not), plus "does not have any mention".since) and judges each answer as of its own date —--rescorereads it from the transcript filename.Re-scored: both models 26/26 on every pass; no previously-ok verdict became wrong in any of the 31 saved runs. New
tests/test_eval_scoring.py; backend 234 passed, mypy clean. Eval tooling only — no app change.🤖 Generated with Claude Code