release: tha-g2p-base-1.0 (gated ByT5-base student, +4.76pp PER)

Sequence-level KD from B-K/umt5 teacher (4.43% PER): 48,757 beam-4
labels, gate student 9.19% (+4.76 <= +5pp). Distill runner hardened
along the way: resumable beam-4 labeling with OOM split-retry,
token-budget batching, byte-length truncation in the training collate
(the OOM saga's root cause), gradient checkpointing, degenerate-label
filtering, teacher recovery script with secryst-recipe stages.

Author: Ronald Tse

models/tha-g2p-base/tha-g2p-base-1.0.README.md

@@ -0,0 +1,22 @@
+# tha-g2p-base-1.0
+
+Thai grapheme-to-phoneme (IPA). Client-tier ByT5-base (580M) student
+distilled from the B-K/umt5-thai-g2p teacher via sequence-level KD:
+48,757 beam-4 teacher-generated labels over the Kaikki + epitran-Wikipedia
+corpus (deduplicated, degenerate outputs filtered).
+
+Gate: student 9.19% PER vs teacher 4.43% on the same harness
+(1,219 Kaikki test sentences, beam-4, corpus-level PER) — +4.76pp,
+inside the +5pp distillation budget (docs/DISTILL-SOURCE-PROMPT.md).
+
+Note on the teacher: the secryst 2.32%-PER umt5 artifacts are
+unrecoverable (transformers 5.15 save drops the untied umt5 lm_head;
+the volume's epitran corpus is tone-less). The 2.32% tier re-enters
+this pipeline when secryst ships repaired artifacts; this model
+distills the best verified teacher available (4.43%).
+
+```python
+from interscript_ml import Model
+model = Model.load("tha-g2p-base-1.0")
+model.translate("สวัสดี")
+```

models/tha-g2p-base/tha-g2p-base-1.0.metadata.yaml

@@ -0,0 +1,32 @@
+format: imf-v1
+id: tha-g2p-base-1.0
+task: g2p
+source_script: Thai
+target: IPA
+tokenizer: bytes
+opset: 14
+decoder: kv
+precision: fp32
+license: BSD-3-Clause
+trained_from: >-
+  sequence-level KD from the B-K/umt5-thai-g2p-v2-0.5k teacher (4.43%
+  PER on this harness; secryst's saved umt5 artifacts are unusable —
+  transformers 5.15 dropped the untied lm_head — and the volume's
+  epitran corpus is tone-less, so the published 2.32% tier is
+  unrecoverable until secryst regenerates it); 48,757 beam-4
+  teacher-generated labels; ByT5-base init google/byt5-base; checkpoint
+  secryst-checkpoints:/secryst_thai_g2p_distill_small/run-004/best
+metrics:
+  - name: per_teacher
+    value: 4.43
+    protocol: >-
+      beam-4, corpus-level PER (total_ed/total_gold over chars of
+      joined-piece decode); 1,219 Kaikki Thai test sentences;
+      B-K/umt5-thai-g2p-v2-0.5k teacher
+    source: interscript/ml-models src/gpu/modal_distill.py::evaluate_per
+  - name: per_student
+    value: 9.19
+    protocol: >-
+      beam-4, corpus-level PER, same harness as the teacher; gate
+      +4.76pp <= +5pp (docs/DISTILL-SOURCE-PROMPT.md); exact match 90.81%
+    source: interscript/ml-models release tha-g2p-base-1.0