feat: Penfield audit corrections for the LoCoMo answer key#17
Merged
Conversation
The April 2026 Penfield Labs audit (github.com/dial481/locomo-audit) found 156 answer-key errors in LoCoMo's 1,540 usable questions: hallucinated facts, temporal arithmetic mistakes, speaker-attribution errors, and wrong evidence citations. Scoring against the original key caps a perfect system at ~93.6% and penalizes correct systems on the broken 6.4%. - datasets/locomo_audit.py: fetch the 10 per-conversation error files at a pinned audit commit (9493fb4b), merge into corrections.json with checksum provenance; loader validates shape and rejects duplicates. - convert locomo --audit-corrections: corrected answers replace expected_answer (QA scoring) and corrected citations replace evidence (retrieval ground truth). Corrected queries carry audit_corrected + audit_error_type metadata. Every correction is cross-checked against the dataset's question text at conversion time so audit/dataset drift fails loudly instead of silently mis-correcting. - justfile recipes (bench-convert-locomo-corrected) and README section. Verified against the real dataset + real audit: all 156 corrections applied (type distribution matches the audit's published counts: 57 WRONG_CITATION, 33 HALLUCINATION, 26 TEMPORAL_ERROR, 24 ATTRIBUTION_ERROR, 13 AMBIGUOUS, 3 INCOMPLETE), 120 answers and 30 ground-truth sets changed, all 1,830 uncorrected queries byte-identical to an uncorrected conversion, zero question-text mismatches. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: Drew Cain <groksrc@gmail.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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Wires the Penfield Labs LoCoMo audit (April 2026) into the harness: 156 answer-key errors across LoCoMo's 1,540 usable questions get corrected answers and corrected evidence citations at conversion time. This is a prerequisite for publishing LoCoMo numbers that withstand scrutiny — the original key is publicly documented as ~6.4% wrong.
Design
9493fb4band merged into onecorrections.jsonwith checksum provenance — runs are reproducible and the upstream can't drift silently.audit_corrected: trueand the audit'serror_typein metadata, so per-correction impact can be analyzed in any run's artifacts.convert locomo --audit-corrections <path>; without the flag, conversion is byte-identical to before (covered by test).Verification
🤖 Generated with Claude Code