version string | description string | pairs list |
|---|---|---|
heldout-v0 | Held-out dx3 recall goldenset (20 pairs), DISJOINT from the 23-pair goldenset v1.0.0. Built for the generalization gate: every pair targets a canon record that was fulltext-confirmed present in the live dx3 instance BEFORE inclusion, so a recall@5 miss means weak retrieval/ranking, not absence. Used to prove the Layer-... | [
{
"id": "core_surfaces_principle",
"question": "What ASIF architecture principle separates the durable core from the surfaces that present it?",
"expected_anchor": "Core and Surfaces Principle",
"existence_probe": "Core and Surfaces Principle CSP",
"fulltext_confirmed": true
},
{
"id": "... |
Dx3 Recall Generalization Benchmark — held-out v0
Author: Asif Waliuddin · NXTG.AI · CC BY 4.0
A held-out recall goldenset (20 query→target pairs), disjoint from the tuning set, built to test whether a retrieval fix generalizes rather than memorizes the set it was tuned against. Every target was full-text-confirmed present in the live store before inclusion — so a recall@5 miss means weak retrieval/ranking, not absence. Honest methodology artifact: a generalization gate, not a leaderboard.
heldout-goldenset-v0.json— the 20 pairs (id, question, target).
Companion research: CRUCIBLE — Measurement Integrity (DOI 10.57967/hf/9867).
- Downloads last month
- 11