eval: record scifact + cmteb public-anchor calibration (Milestone B)#6
Closed
helebest wants to merge 1 commit into
Closed
eval: record scifact + cmteb public-anchor calibration (Milestone B)#6helebest wants to merge 1 commit into
helebest wants to merge 1 commit into
Conversation
First real-vector baselines, validating the eval chain end-to-end on non-saturated, literature-anchored data (the Phase 0→1 advance criterion). - scifact (en): doc/hybrid ndcg_at_10 0.689 — matches BEIR literature 0.67 (Δ 0.019, within ±0.10) and clears dikw-core's committed floor. RRF lift real (bm25 0.651 < vector 0.673 < hybrid 0.689). - cmteb-t2-subset (zh): reproduces dikw-core's calibrated subset baseline within noise (bm25 ndcg_at_10 0.840 exact; hybrid 0.946 vs 0.952; vector 0.943 vs 0.942); clears all dataset thresholds. jieba CJK confirmed (0.840, not the degenerate 0.031). NB: this is a curated 300-q subset, not the full CMTEB leaderboard (~0.50) — see its dataset.yaml. - Cross-cutting: multi-batch embedding confirmed (5183 / 5000 chunks ≫ 16/batch, the Phase 0 gap); read-only held (scifact materialized into gitignored paths, dataset.yaml restored, dikw-core tree clean). No gates set — calibration only; per-language thresholds wait for the in-house domain-bilingual-v1 / negatives-ood-v1 sets. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
helebest
force-pushed
the
ci/lint-and-eval-gates
branch
from
72527bd to
6d62374
Compare
helebest
force-pushed
the
eval/anchor-calibration
branch
from
8cbc031 to
1af56be
Compare
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What
First real-vector eval baselines, recorded in
reports/BASELINES.md. Validates the eval chain end-to-end on non-saturated, literature-anchored data — the documented Phase 0→1 advance criterion (docs/dikw-eval-plan.md§3). This is Milestone B, following the CI floor in #5.Results (doc/hybrid canonical,
--retrieval all, 1 run each)ndcg_at_10 ≈ 0.67(Δ 0.019, within ±0.10) and clears dikw-core's committed floor. RRF lift real: bm25 0.651 < vector 0.673 < hybrid 0.689.ndcg_at_10 0.840exact; hybrid 0.946 vs 0.952; vector 0.943 vs 0.942). This is a curated 300-query subset (distractor-padded), not the full CMTEB leaderboard (~0.50) — see itsdataset.yaml. jieba CJK confirmed (0.840, not the degenerate unicode61 0.031).Cross-cutting checks
corpus/+queries.yamlmaterialized into gitignored paths; its trackeddataset.yamlbacked up and restored;git -C dikw-core statusstays clean.Not in this PR
No gates set — calibration only. Per-language thresholds get set at
observed − marginonce the in-housedomain-bilingual-v1/negatives-ood-v1sets exist (rest of Phase 1).🤖 Generated with Claude Code