From 7ee641bde595506bd675121847d3c9a19f0cbfc1 Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Mon, 31 Aug 2026 09:59:48 +0200 Subject: [PATCH] =?UTF-8?q?docs:=206ep=20subset=202.0062=20=E2=80=94=20res?= =?UTF-8?q?idual=20was=20undertraining,=20not=20domain?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- docs/RESULTS.md | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/docs/RESULTS.md b/docs/RESULTS.md index 000b7de..6500740 100644 --- a/docs/RESULTS.md +++ b/docs/RESULTS.md @@ -423,3 +423,15 @@ The law sharpens: the pretrained WIDTH is load-bearing - projection destroys the representation at any compression; depth is the compressible axis. First verdict carrying the label-provenance hash (labels sha256 e70ce991..., 137.2MB recorded in final_eval.json). + +## ara-diac-small-2-6ep — doubling epochs collapses the residual (2026-08-31, subset) + +Same as the 2.0 rung (r7 labels, Muon, vanilla ByT5-small) at 6 epochs +instead of 3 (21,990 steps, final CE 0.0013). 300-subset verdict: +teacher in-run 1.2887, student **2.0062** — gate_delta 0.72, within +striking distance of the strict teacher+0.5 gate, vs 3.66 (1.0) and +3.81 (layerdrop) on the same subset. **This overturns the E2/E3 +residual attribution**: the 2.25pp residual was mostly optimization +(undertraining), not domain coverage. Full-set gate running - given +three prior subset-overstatement instances, the honest number waits +there. Provenance: labels sha256 e70ce991 (137.2MB) recorded.