Judge calibration: 40 human labels, weighted kappa 0.53 adequacy / 0.27 fluency / 0.21 localization#7
Merged
Merged
Conversation
…3/0.27/0.21) 40 blind human labels over the seeded stratified sample. Quadratic-weighted Cohen's kappa vs the qwen3.5 judge: adequacy 0.526 (moderate - usable), fluency 0.267 and localization 0.213 (weak - judge systematically more lenient than a Taiwanese reader on terminology). Docs updated so quality claims lean on chrF + adequacy and treat judge localization as upper bound. Co-Authored-By: Claude Fable 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Closes the calibration loop from #6: 40 blind human ratings over the seeded stratified sample (
eval/results/labeling.html), computed bypnpm bench:agreement.Interpretation shipped into
docs/BENCHMARK.md/ README: the qwen3.5 judge is moderately trustworthy on adequacy, weak on fluency and Taiwan-localization (systematically more lenient than a Taiwanese reader), so the report leans on chrF + adequacy and reads the judge localization column as an upper bound. Calibration told us which parts of the judge to trust - that is the point of the exercise.eval/dataset/human-labels.jsoncommitted for reproducibility;eval/AGREEMENT.mdgenerated.🤖 Generated with Claude Code