feat(WO68/B7): khm-dict corpus pinned (Chuon Nath lexicon phase 1)

TODO.sota/68-b7-khmer-lexicon.md

@@ -1,6 +1,13 @@
 # WO68 — B7: Khmer — Chuon Nath lexicon conditioning (CER < 15)
 
-Status: PLANNED · launchable now (data on disk: khmer-dict-spice)
+Status: IN_PROGRESS (phase 1 landed 2026-10-21; phase 2 = training arm)
+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
+(size-verified). License LGPL-2.1 recorded; owner note on
+LGPL-for-data semantics. NEXT: hit-rate vs the khm-latn eval set (in
+the phase-2 job, which mounts it), lexicon-conditioned ByT5 fallback
+training, three-way probe.
 Goal: khm-latn 27.4 → CER < 15 with hit-rate reported.
 Method: build the Chuon Nath-conditioned dataset (headword→reading
 pairs, syllabification-aware); lexicon-conditioned ByT5 (retrieval or

corpora/khm-dict.yaml

@@ -0,0 +1,19 @@
+id: khm-dict
+task: transliteration
+sources:
+  # Chuon Nath dictionary data, digitalized by the Open Institute
+  # (SPICE / USAID-funded) — github.com/interscript/khmer-dict-spice.
+  # Repo LICENSE: LGPL-2.1 (SPDX-open; gate clears). Owner note:
+  # LGPL semantics for a data corpus are unusual; flagged for owner
+  # review without blocking (registry law: SPDX-open clears).
+  - kind: bucket
+    id: Interscript/isx-training/run-080-khm-lexicon
+    revision: run-080-khm-lexicon
+    license: LGPL-2.1
+    mount: /data-khm
+splits:
+  # Gold reading pairs (word -> IPA, Chuon Nath via SPICE ipa_dic):
+  # 3261 pairs, 326 held out as test. SBBIC vocabulary (72,048 words,
+  # no readings) carries the lexicon-coverage layer.
+  train: {path: khm-lexicon-train.jsonl, rows: 74983, sha256: 83eba4945044543c6a0af467ecdcb72a16b2f552fa86e041bb5eeca6e9b2e9be}
+  test: {path: khm-lexicon-test.jsonl, rows: 326, sha256: 23385241b606f3ca1f74d09df5bf34fe32b364545171edf157f0c90a58bd5b18}