From c58195c863946c92c08856eb09196ffed97a5566 Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Fri, 23 Oct 2026 12:39:10 +0800 Subject: [PATCH] =?UTF-8?q?docs(WO67):=20v5=20POSITIVE=20=E2=80=94=20PER?= =?UTF-8?q?=2020.25=20beats=20v1's=2022.56;=20length-prior=20recipe=20conf?= =?UTF-8?q?irmed?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- TODO.sota/67-b6-thai-lexicon.md | 14 +++++++++++--- 1 file changed, 11 insertions(+), 3 deletions(-) diff --git a/TODO.sota/67-b6-thai-lexicon.md b/TODO.sota/67-b6-thai-lexicon.md index f8ce3367..7b9e2dbe 100644 --- a/TODO.sota/67-b6-thai-lexicon.md +++ b/TODO.sota/67-b6-thai-lexicon.md @@ -35,9 +35,17 @@ shows only after the schedule is fixed. 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. +**V5 VERDICT (2026-10-23): POSITIVE — recipe found.** v5 (92,669 +length-filtered rows, BS4 x 2ep) scores **PER 20.25%** on the v1 +split — BEATS the v1 reference (22.56) with 4.1x more training types +and honest generalization, while v4-final (37.4, unfiltered, same +schedule) proves the schedule alone never fixed it. The +length-prior mechanism is confirmed end to end. NEXT: (a) rebuild the +eval set without its 15 defect rows and re-score everything; (b) +convergence arm — v5 recipe, 4-6 epochs, optionally re-add filtered +compounds in a final curriculum phase for long-word coverage; (c) the +converged model becomes the fallback-learner release candidate +(fixtures + gates). 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