From efb9df3e38264dd121e9dc797cfe97ff44d376ae Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Sat, 31 Oct 2026 02:08:33 +0800 Subject: [PATCH] =?UTF-8?q?docs:=20WO68=20=E2=80=94=20attempt=203=20nomina?= =?UTF-8?q?l=20EXPANSION=5FPAYS=20but=20statistically=20a=20tie=20(27.08?= =?UTF-8?q?=20vs=2027.42)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- TODO.sota/68-b7-khmer-lexicon.md | 20 ++++++++++++++++++++ 1 file changed, 20 insertions(+) diff --git a/TODO.sota/68-b7-khmer-lexicon.md b/TODO.sota/68-b7-khmer-lexicon.md index ac3130c..726cbfd 100644 --- a/TODO.sota/68-b7-khmer-lexicon.md +++ b/TODO.sota/68-b7-khmer-lexicon.md @@ -166,3 +166,23 @@ train anchor, exact run-081 split) -> gate: student CER < teacher CER on the identical 895 test rows (the lineage comparison vs 27.4). Verdict script: train_khm_expansion3.py (stage-b3 outputs under bucket khmer-latn/). + + +**Attempt 3 verdict: nominal EXPANSION_PAYS, statistically a TIE** (2026-10-31, +run-082f). All 74,992 unique SBBIC words labeled with the shipped +khm-latn-1.0 (0 empties this pass). Student (byt5-small, 74,992-convention +weak labels + 16,120 translit anchor) vs teacher on the identical 895-row +translit test split: + +- student CER 27.079% / EM 59.89% +- teacher CER 27.417% / EM 59.66% (reproduces the published 27.42 + benchmark exactly — protocol identity verified) + +The +0.34 CER delta is within noise at n=895. Interpretation: weak-label +expansion REPLICATES the teacher (the labels are the teacher's own +outputs — training converges to the teacher, including its errors). No +evidence expansion alone reaches the CER<15 B7 target. The line's next +real lever is NEW GOLD, not more teacher output: SEAlang / Chuon Nath +full IPA remain the owner-held data asks (B7 blocker unchanged), since +a second decorrelated latn-convention teacher for agreement-denoising +does not exist today.