increase relevance cutoff during model calibration

This commit is contained in:
Bogdan Abaev 2026-08-26 10:56:49 -07:00
parent 615d395bf8
commit 21036418e4

View file

@ -2946,10 +2946,10 @@ Zotero.Embeddings.Indexing = new function () {
Zotero.Embeddings.Calibration = new function () {
// Where the score floor goes, as a percentile of the null distribution --
// what this model scores between texts with nothing to do with each other
// (`mismatched` in measure()). At 0.99 only the top 1% of unrelated pairs
// (`mismatched` in measure()). At 0.999 only the top 0.1 of unrelated pairs
// reach it; raising it cuts more noise and more weak-but-real matches along
// with it, lowering it keeps both.
const NULL_PERCENTILE = 0.99;
const NULL_PERCENTILE = 0.999;
// Where the Relevance bar fills, as a percentile of the other distribution
// measure() collects -- `matched`, each query against its own passage. At
// 0.5 a full bar means "as good as this model's typical real match".