From 51a911b4253745138e1ae297507fc97bf076b0eb Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Thu, 22 Oct 2026 21:03:27 +0800 Subject: [PATCH] =?UTF-8?q?docs(WO68):=20khm=20Chuon=20Nath=20arm=20traine?= =?UTF-8?q?d=20=E2=80=94=20dev=20CER=2015.02%=20on=202,935=20pairs;=20supp?= =?UTF-8?q?ly=20wall=20documented?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- TODO.sota/68-b7-khmer-lexicon.md | 13 ++++++++++++- 1 file changed, 12 insertions(+), 1 deletion(-) diff --git a/TODO.sota/68-b7-khmer-lexicon.md b/TODO.sota/68-b7-khmer-lexicon.md index 5b04835a..1d0e182c 100644 --- a/TODO.sota/68-b7-khmer-lexicon.md +++ b/TODO.sota/68-b7-khmer-lexicon.md @@ -1,6 +1,17 @@ # WO68 — B7: Khmer — Chuon Nath lexicon conditioning (CER < 15) -Status: IN_PROGRESS (phase 1 landed 2026-10-21; phase 2 = training arm launch) +Status: IN_PROGRESS — Chuon Nath reading arm TRAINED (2026-10-22, +run-080-khm-lex-2): ByT5-small on 2,935 gold pairs -> dev CER 15.02% +(326 held-out dictionary rows, 8 epochs, descending to the end). +Caveats: dictionary-internal dev (memorization scale), NOT novel-word +performance; SBBIC's 72,048 staged rows are word-only (ipa: null) — +**the open Khmer phoneme supply is exhausted at ~3.3k gold pairs** +(ipa_dic.txt 3,286 lines = the whole local source). The CER < 15 +khm-latn target therefore runs through (a) the khm-latn lineage +recovery (Khmer->Latin pairs, different and larger supply) or (b) +owner-gated licensing (SEAlang / full Chuon Nath IPA). GPU finding: +ByT5-small + 2.9k pairs trains to sub-16% in ~7 min — the learner is +never the bottleneck here. Phase 1 DONE: khm-dict corpus built + pinned (corpora/khm-dict.yaml) — 3,261 gold Chuon-Nath (word→IPA) pairs, 326-row holdout, 72,048 SBBIC vocab rows; staged at bucket Interscript/isx-training/run-080-khm-lexicon