Repository navigation
bench: answer keys and outcome scorer for four rival rows - #1966
Closed
remigiusz-antczak wants to merge 1 commit into
Closed
remigiusz-antczak wants to merge 1 commit into
remigiusz-antczak wants to merge 1 commit into
Conversation
…rows Add keys.json (3-6 must-hit and 1-3 must-not regex items per scenario) for the 40 held-out scenarios of business-ops, growth-analytics, idea-critic and contribution. Keys test what an answer adds beyond its prompt (echo baseline max 0.25), accept number forms and paraphrases, and carry no Perun verdict tokens. rival_eval.py gains key_score (per-arm hit rate and violation rate, no judge), an answer_key block in results.json, a keys subcommand that rescores cached outputs offline and fails on a missing arm directory, and key hygiene in check. Must-nots are negation-aware and skip placeholders; contribution leak and force-push checks scan only the drafted message, and a missing draft is flagged no_draft. No paid call. Tests: 61/61 (test_rival_answer_keys 14, test_rival_eval 47) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014QNFFPWVcPUbeuvwM5fk2Z
Collaborator
Author
|
Superseded by #1983. |
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Adds answer keys and a deterministic hit-rate scorer for four rival-bench rows (business-ops, growth-analytics, idea-critic, contribution). These rows can then be scored on outcomes instead of judge preference.
rival_eval.py keysscores each arm's outputs. It reports per-scenario hits, the must-not violations, andno_draftwhere no drafted message was found. It fails if an arm directory is missing.Fairness guards, after the independent review:
Not yet done: calibrating the keys against real arm outputs (no cached outputs match the current ids). The first paid run should spot-check keys before any row verdict counts.
Tests: 61/61 (answer keys 14, rival_eval 47), run on the commit's own tree. Independent review: 0🔴; the 8🟡 and 1🔵 are fixed here.
🤖 Generated with Claude Code
https://claude.ai/code/session_014QNFFPWVcPUbeuvwM5fk2Z