From 791fa40c637f2f17b8d0949f468ddcb615e07a52 Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Fri, 23 Oct 2026 10:27:55 +0800 Subject: [PATCH] =?UTF-8?q?docs(WO67):=20five-finding=20diagnostic=20chain?= =?UTF-8?q?=20=E2=80=94=20length=20prior=20identified;=20v5=20is=20the=20c?= =?UTF-8?q?orrective=20arm?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- TODO.sota/67-b6-thai-lexicon.md | 20 ++++++++++++++++++++ 1 file changed, 20 insertions(+) diff --git a/TODO.sota/67-b6-thai-lexicon.md b/TODO.sota/67-b6-thai-lexicon.md index 400841ea..f8ce3367 100644 --- a/TODO.sota/67-b6-thai-lexicon.md +++ b/TODO.sota/67-b6-thai-lexicon.md @@ -18,6 +18,26 @@ big-data arms degrade identically -> the schedule is the confound, not the corpora. run-080-thai-lex-v4 (v3 data, BS4 x 2ep ~= 65k updates, matching v1's budget) is the control; wiki tier's own value shows only after the schedule is fixed. + +**Diagnostic chain (2026-10-23, five verified findings):** +1. Engine drift = 0.0000 PER — royin 5.3.5 output is byte-identical to + the stored v1 targets on clean rows; convention is NOT the issue. +2. v1-test carries 15 unfiltered old-engine defect rows (1.5% of gold + tokens; e.g. 'd u ŋ' x8) — small, but the eval set needs a clean + rebuild. +3. v1's split was clean (0 overlap) — its 22.56 is honest + in-distribution generalization on orst-only. +4. v2/v3/v4 TRAIN on 790 of the v1-test words (24.5%) and still lose — + unseen-word PER is worse still (~50% est.). +5. Mechanism: the expanded corpora shifted the output-LENGTH prior + (volubilis avg 25.6 phoneme tokens vs test 10.4) — long-compound + dominance wrecks short-word generation. v4-ep0 at 41.8 PER also + rules out update-budget alone. +**v5 (run-080-thai-lex-v5)** = the corrective: 92,669 rows filtered to +<=20 phoneme tokens (deployment-length distribution), BS4 x 2ep. +Decision rule: v5 ~<= 22.56 => recipe found (then scale epochs, re-add +filtered compounds for coverage); v5 still ~39 => the mechanism is +wrong and the next suspect is per-source training curricula. RTGS encodings landed maps#204; IPA layers = WO82. Lexicon phase DONE: corpora/tha-phonemic.yaml -- 28,644 train / 3,225 test (word -> phonemic-with-tones pairs, ORST vocabulary x pythainlp