From 966ae859d5ade509f50b05b243238906b2ccd177 Mon Sep 17 00:00:00 2001
From: Ronald Tse <ronald.tse@gmail.com>
Date: Wed, 28 Oct 2026 00:10:50 +0800
Subject: [PATCH] =?UTF-8?q?feat(WO88):=20khm-phonemic-1.0=20release=20cand?=
 =?UTF-8?q?idate=20=E2=80=94=20fixture=20+=20entry=20(parity=20GREEN,=2098?=
 =?UTF-8?q?.67%=20row-exact)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 conformance/khm-phonemic-vectors.json | 58 +++++++++++++++++++++++++++
 models.yaml                           | 28 +++++++++++++
 2 files changed, 86 insertions(+)
 create mode 100644 conformance/khm-phonemic-vectors.json

diff --git a/conformance/khm-phonemic-vectors.json b/conformance/khm-phonemic-vectors.json
new file mode 100644
index 00000000..5be46e1e
--- /dev/null
+++ b/conformance/khm-phonemic-vectors.json
@@ -0,0 +1,58 @@
+{
+ "model": "khm-phonemic-1.0",
+ "index_version": 12,
+ "decode": {
+  "mode": "single-row-greedy"
+ },
+ "source": "run-090-khm-multi multi-task (secryst lineage + Chuon Nath IPA); WO88",
+ "vectors": [
+  {
+   "src": "ភាសាវិទូ",
+   "out": "pheasaeavitou"
+  },
+  {
+   "src": "គ្របគ្រង",
+   "out": "krobakrong"
+  },
+  {
+   "src": "រាជអាណាចក្រ",
+   "out": "anachakr"
+  },
+  {
+   "src": "លសិកា",
+   "out": "lseka"
+  },
+  {
+   "src": "កំសាក",
+   "out": "kamsaak"
+  },
+  {
+   "src": "ឱត្តប្បៈ",
+   "out": "aottabb"
+  },
+  {
+   "src": "រណៈ",
+   "out": "rn"
+  },
+  {
+   "src": "ដំរង់",
+   "out": "damrng"
+  },
+  {
+   "src": "ខ្ចីឲ្យ",
+   "out": "khchei aoy"
+  },
+  {
+   "src": "បាល័ដ្ឋ",
+   "out": "baldth baldtha"
+  },
+  {
+   "src": "មានទោស",
+   "out": "meantosa"
+  },
+  {
+   "src": "បិណ្ឌមួយ",
+   "out": "bendamuoy"
+  }
+ ]
+}
\ No newline at end of file
diff --git a/models.yaml b/models.yaml
index 4cd90fbe..ce34ec36 100644
--- a/models.yaml
+++ b/models.yaml
@@ -36,6 +36,34 @@ renames:
   urd-g2p-base-1.0: urd-ipa-base-1.0
 
 models:
+  khm-phonemic-1.0:
+    task: g2p
+    precision: fp32
+    filename: khm-phonemic-1.0.zip
+    url: https://github.com/interscript/interscript-models/releases/download/UNPUBLISHED/khm-phonemic-1.0.zip
+    sha256: b1eb0cd613fbd06a33f4ec15902f6e425469be89606f629f5e10bd45ac6d4283
+    size: 1112390964
+    metrics:
+    - name: cer
+      value: 24.50
+      protocol: dev CER, multi-task test split (997 rows), single-row greedy
+      source: run-090-khm-multi oepoch-2 eval (WO88; like-for-like vs legacy 27.57%)
+    - name: cer_legacy_split
+      value: 24.16
+      protocol: dev CER after the accidental resume-extend (oepoch-7)
+      source: run-090-khm-multi oepoch-7 eval
+    parity:
+      samples: 600
+      cer_delta: 0.0032
+      row_exact_match: 0.9867
+      note: >-
+        torch vs shipped zip, single-row greedy, 600 secryst-split rows
+        (run-090-khm-parity, 2026-10-27)
+    license: BSD-3-Clause
+    trained_from: >-
+      ByT5-small multi-task <latn>+<ipa> prefixes; corpora/khm-latn.yaml
+      (17,015 secryst-lineage pairs) + khm-lexicon-train (2,935 Chuon Nath
+      IPA); 4 epochs BS4; WO88
   tha-phonemic-1.0:
     task: g2p
     precision: fp32