Repository navigation
docs: align claims with committed evaluation artifacts - #9
Merged
Merged
Conversation
- README: disclose the three v3 runs, their differing results and why run-03 is reported. - docs/llm-validation.md, evaluation/v2/README.md: run-03 is the current record; run-01 is historical. - Say precisely which failure mode the extractive contract removes and which remain. - evaluation/v3/README.md: use "previously exposed" for the 42 questions; document what changed between runs. - RELEASE_NOTES.md: use the committed benchmark and retrieval results; mark uncommitted earlier figures as historical.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Documentation-only. No code, results, manifests or run records are changed.
evaluation/v3/run-0{1,2,3}/summary.json: false accepts / covered / false rejects 12/18/3, 8/23/3, 12/21/2), why run-03 is the reported one, and that run-02 is the best of the three.docs/llm-validation.md,evaluation/v2/README.md: run-03 is the current record; run-01 is historical (its manual review stays linked fromevaluation/v3/README.mdas historical).docs/llm-validation.md: the extractive contract removes model-authored prose; irrelevant selection and false accepts remain.evaluation/v3/README.md: the 42 questions are "previously exposed" (as in the manifests'limitationfield); adds a section on what did and did not change between runs, from the manifests and commit history. The generatedrun-*/report.mdfiles are not edited; a note explains their "held-out" label.RELEASE_NOTES.md: benchmark now citesevaluation/v3/benchmark-local-http.json(p50 0.857 / p95 0.920 / max 2.110 ms); retrieval figures now come fromevaluation/results_modes.json(all 30 and the 20-question subset, all modes). The earlier 0.448/0.566 ms, 0.9375 and nomic 0.9167/0.8667 figures had no committed output and were measured on the earlier 20-question set (commit 2ecd148 and before); they are kept only as marked historical figures.Checks run locally (same as
.github/workflows/ci.yml): cargo fmt --check, cargo check, cargo clippy -D warnings, cargo test (37 tests), evaluation/v3 unit tests, evaluation/evaluate.py, evaluation/evaluate_modes.py (its in-place output was discarded).