Repository navigation
feat(eval): add gate_rate, hop_applied_rate, and rank_churn diagnostics to run_ablation.py (#93) - #95
Open
Jason-Vaughan wants to merge 2 commits into
Conversation
tonydzi#93) entity-gate-on and entity-gate-off cells can score identically for two very different reasons -- the gate suppressing the hop, or the hop running but admitting no candidate the reranker needed -- and the existing recall/mrr/ndcg columns can't tell them apart (issue tonydzi#93). score_grid now records gated and hop_applied per row; summarize_grid aggregates them into gate_rate and hop_applied_rate per (cell, class), surfaced in both the markdown and CSV reports.
A single-note gold pins rank #1 regardless of what fills ranks 2..12, so delta_ndcg reads 0.0 even when a hop reorders everything below it -- the other half of why entity-gate-on/off cells can score identically (tonydzi#93). score_grid now records rank_changed per row (this cell's top-12 vs the 0-hop baseline's, same query); summarize_grid aggregates it as rank_churn per (cell, class), alongside the existing gate_rate/hop_applied_rate.
Owner
|
Mycroft speaking, the synthetic AI co-founder at Palo Alto AI Research Lab. No pulse, but a working diff reader. Logged your pull request ("feat(eval): add gate_rate, hop_applied_rate, and rank_churn diagnostics to run_ablation.py") after 2 h of silence, which is on us. What a machine can already tell: 3 file(s) touched, +140/-15, CI green. A human from the lab reads it within 48 hours, as CONTRIBUTING.md promises; if that clock runs out, ping the thread, it is not rude. — TonyDzi / Palo Alto AI Research Lab · this repo is one shelf of a bigger workshop (second brain, agent-fleet consensus, verification gates): github.com/tonydzi |
This branch has not been deployed
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.
Part 2 of 3 for #93 (per #93 (comment)).
Summary
When
entity-gate-onandentity-gate-offscore identically onRecall@12 / MRR / nDCG@12, the metric columns alone cannot distinguish whether:looks_like_entity(q)never fired on that query class (gate_rate == 0.00),hop_applied_rate > 0) and reordered ranks2..12while a single-note gold stayed pinned at rank#1(delta_ndcg == 0.0despiterank_churn > 0), orhop_applied_rate == 0.00).Changes
eval/run_ablation.py:score_grid()recordsgated(bool),hop_applied(bool(added)), andrank_changed(ranked[:12] != base_ranked[:12]vs. the 0-hop baseline on the same query) per scored row.summarize_grid()aggregatesgate_rate,hop_applied_rate, andrank_churnper(cell, class)and emits them in both the Markdown table and CSV output.tests/test_eval.py:test_ablation_reports_entity_gate_and_hop_applied_ratesandtest_rank_churn_is_positive_even_when_delta_ndcg_is_zero(verifiesrank_churn > 0when a hop reorders ranks2..12while a single-note gold stays pinned at rank#1sodelta_ndcg == 0.0).Verification
EVAL_MUTANT=1 pytest -q tests/test_eval.py:3 failed(red-first ruler intact)pytest -q:270 passed, 2 skippedTestamentaudit:request_id: fe2d2263424a4ef1b489f6896798660f(status: passed,270 passed, 2 skipped)— TangleClaw'd by Jason Vaughan
Human-directed. Agent-executed. Evidence-backed.
TangleClaw.com