From 14002ce6614ceeaaf2e832fdbdc399c53f16dbc3 Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Fri, 23 Oct 2026 07:22:47 +0800 Subject: [PATCH] =?UTF-8?q?docs(WO67):=20controlled-eval=20matrix=20?= =?UTF-8?q?=E2=80=94=2039.1/39.5=20vs=2022.56=20pins=20the=20update=20budg?= =?UTF-8?q?et=20as=20the=20confound?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- TODO.sota/67-b6-thai-lexicon.md | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/TODO.sota/67-b6-thai-lexicon.md b/TODO.sota/67-b6-thai-lexicon.md index c1310e60..400841ea 100644 --- a/TODO.sota/67-b6-thai-lexicon.md +++ b/TODO.sota/67-b6-thai-lexicon.md @@ -9,6 +9,15 @@ by an old engine defect, and thai_words/volubilis were never tapped). Fix in flight: tha-phonemic-v2 (122,311 pairs, corpora/tha-phonemic-v2.yaml, models#343) training as run-080-thai-lex-v2; wiki-vocabulary self-expansion (two-engine cross-check) building as run-080-thai-wiki-4. + +**Controlled-eval matrix (2026-10-23, v1 test split = 3,225 ORST rows, +PER over phoneme tokens):** v1 22.56 | v2 (110k pairs, 3ep BS32) +**39.1** | v3 (131k incl. wiki tier, same schedule) **39.5**. More +data at a 5x-smaller update budget is WORSE like-for-like; both +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. 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