feat(tha): tha-g2p-small-1.0 client tier — run-003 byt5-small registration

Re-measured run-003 on the beam-4 harness (12.06% PER, EM 87.94%,
+7.63pp over the 4.43% teacher, n=1219). Capacity-limited but the
smallest non-collapsing rung; exported at int8 (~300MB) per the
frontier table. Frontier row updated 12.63 -> 12.06 (full-label run).

Author: Ronald Tse

docs/RESULTS.md

@@ -60,6 +60,26 @@ IMF test split (1,864 long sentences), same harness for both models
 Shrink cost +5.58pp — inside the ~5.6pp budget pre-accepted for this
 pair (rababa docs/DISTILL-SOURCE-PROMPT.md section 2).
 
+## tha-g2p-small-1.0 — Thai G2P client tier (2026-08-22)
+
+The client-tier release of the Thai G2P distillation: run-003,
+ByT5-small student on the full label set (48,757 usable beam-4 labels
+from the B-K/umt5-thai-g2p-v2-0.5k teacher). Same harness as
+tha-g2p-base-1.0 (beam-4, corpus-level PER, 1,219 held-out Kaikki Thai
+test sentences, `src/gpu/modal_distill.py::evaluate_per`; checkpoint
+re-measured 2026-08-22 for this release).
+
+| Model | PER | Exact match |
+|---|---|---|
+| Teacher (B-K/umt5 hub base) | 4.43% | 95.57% |
+| **Student (ByT5-small, client rung)** | **12.06%** | 87.94% |
+
+Shrink cost +7.63pp — outside the +5pp server-tier gate (that gate is
+met by tha-g2p-base-1.0 at 9.19%): shipped anyway per the frontier
+below, as the smallest artifact that does not collapse. Exported at
+int8 (~300MB); see the frontier table for why no smaller rung exists
+today.
+
 ## Client-tier size–quality frontier (2026-08-22)
 
 Thai G2P, same harness (beam-4 corpus PER, 1,219 Kaikki sentences; teacher
@@ -70,7 +90,7 @@ B-K umt5 4.43%):
 | custom 8+8 d384 | random | 33M | ~30MB | 75.80 (collapsed) |
 | custom 8+8 d384 + bridges | random | 33M | ~30MB | 71.12 |
 | custom 10+10 d512 + bridges | random | 70M | ~70MB | 78.51 |
-| ByT5-small | pretrained | 300M | ~300MB | 12.63 |
+| ByT5-small | pretrained | 300M | ~300MB | 12.06 |
 | ByT5-base (server tier) | pretrained | 580M | 1.2GB fp32 | 9.19 |
 
 Findings: (1) random-init byte-level seq2seq collapses regardless of
@@ -79,7 +99,7 @@ capacity at this scale — the microkimi bridges improve structure (75.8 →
 not help (70M = 78.5). (2) ByT5-small's width (d=1472) dominates its
 parameter count — depth-pruning yields no useful intermediate rung
 (263M). (3) The pretrained rung is the whole quality cliff: 300M at
-12.63% vs 70M at 78.5%.
+12.06% (run-003, full labels; 12.63% on the 23K subset) vs 70M at 78.5%.
 
 Conclusion: G2P client tier ships at the ByT5-small rung (~300MB int8)
 today; a 30–70MB G2P tier requires byte-level pretraining of the small

models/metrics-sources.yaml

@@ -67,3 +67,12 @@ heb-diac-small-1.0:
   tables:
     - {row: "Teacher (s43, ByT5-base)", column: DER, as: der_teacher_greedy}
     - {row: "Student (ByT5-small, gate)", column: DER, as: der_student_greedy}
+tha-g2p-small-1.0:
+  repo: interscript/interscript-ml
+  ref: release/tha-g2p-small-1.0
+  path: docs/RESULTS.md
+  anchor: tha-g2p-small-10-thai-g2p-client-tier-2026-08-22
+  protocol: "beam-4, corpus-level PER (total_ed/total_gold over chars of joined-piece decode); 1,219 held-out Kaikki Thai test sentences"
+  tables:
+    - {row: "Teacher (B-K/umt5 hub base)", column: PER, as: per_teacher}
+    - {row: "Student (ByT5-small, client rung)", column: PER, as: per_student}

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

@@ -1,14 +1,16 @@
 # tha-g2p-small-1.0
 
 Thai grapheme-to-phoneme (IPA). Client-tier ByT5-small (300M) student
-distilled from the secryst umt5 Thai teacher (2.32% PER, public baseline
-6.37%) via sequence-level KD: the teacher generated 23,295 labels with
-its own sentencepiece tokenizer and the student trained CE on them with
-the canonical byte table.
+for the int8 release (~300MB): sequence-level KD from the verified
+B-K/umt5-thai-g2p-v2-0.5k teacher over 48,757 beam-4 labels, trained
+with the canonical byte table.
 
-First model of the distillation campaign
-(docs/DISTILL-SOURCE-PROMPT.md); identical IMF v1 contract to the
-server-tier models — dynamic fetch, sha256-verified, KV decode.
+Capacity-limited at +7.63pp over the teacher (12.06% vs 4.43% PER) —
+inside the +5pp server-tier gate is tha-g2p-base-1.0; this is the
+smallest artifact that does not collapse (from-scratch 33M/70M students
+and pruned rungs all fail — see the frontier table in
+docs/RESULTS.md). Identical IMF v1 contract to every other model:
+dynamic fetch, sha256-verified, KV decode.
 
 ```python
 from interscript_ml import Model

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

@@ -9,22 +9,27 @@ decoder: kv
 precision: fp32
 license: BSD-3-Clause
 trained_from: >-
-  distilled from the recovered Thai umt5 teacher (transformers 5.14.1
-  re-finetune of B-K/umt5-thai-g2p-v2-0.5k on the 60K Kaikki+epitran
-  corpus; the secryst-saved umt5 artifacts are unusable — 5.15 dropped
-  the untied lm_head) via sequence-level KD; ByT5-small init
-  google/byt5-small; checkpoint
-  secryst-checkpoints:/secryst_thai_g2p_distill_small/run-002/best
+  sequence-level KD from the B-K/umt5-thai-g2p-v2-0.5k teacher (4.43%
+  PER on this harness; the secryst-saved umt5 artifacts are unusable —
+  transformers 5.15 dropped the untied lm_head); 48,757 usable beam-4
+  teacher-generated labels (51,321 raw, deduplicated, degenerate outputs
+  filtered); ByT5-small init google/byt5-small; checkpoint
+  secryst-checkpoints:/secryst_thai_g2p_distill_small/run-003/best.
+  Client tier: capacity-limited at +7.63pp over the teacher — the +5pp
+  gate is met by tha-g2p-base-1.0; this is the smallest rung that does
+  not collapse (docs/RESULTS.md frontier table, ~300MB int8).
 metrics:
   - name: per_teacher
-    value: 2.32
+    value: 4.43
     protocol: >-
-      greedy decode; 1,219 Kaikki Thai test sentences; umt5 teacher,
-      sentencepiece tokenizer; secryst RESULTS.md protocol
-    source: secryst/docs/RESULTS.md#thai-g2p
+      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; exact match 95.57%
+    source: interscript/interscript-ml docs/RESULTS.md#tha-g2p-small-1.0
   - name: per_student
-    value: PLACEHOLDER_STUDENT_PER
+    value: 12.06
     protocol: >-
-      greedy decode; same 1,219 test sentences, same harness as the
-      teacher (interscript/ml-models src/gpu/modal_distill.py::evaluate_per)
-    source: interscript/ml-models release tha-g2p-small-1.0
+      beam-4, corpus-level PER, same harness as the teacher (src/gpu/
+      modal_distill.py::evaluate_per, run-003 checkpoint re-measured
+      2026-08-22); exact match 87.94%; +7.63pp client-tier shrink cost
+    source: interscript/interscript-ml docs/RESULTS.md#tha-g2p-small-1.0

src/gpu/modal_distill.py

@@ -83,6 +83,23 @@
         "mode": "sequence",  # cross-tokenizer: teacher generates, student trains CE
         "note": "umt5 (sentencepiece) teacher -> ByT5-small byte student; +5pp PER gate",
     },
+    "tha-g2p-client": {
+        # the shipped client rung: run-003, ByT5-small on the full label
+        # set — capacity-limited (+7.6pp) but the smallest artifact that
+        # does not collapse (see docs/RESULTS.md frontier table)
+        "teacher": "B-K/umt5-thai-g2p-v2-0.5k",
+        "teacher_is_hub": "true",
+        "teacher_volume": "secryst",
+        "student_init": "google/byt5-small",
+        "train": "thai-ipa-expanded/train.jsonl",
+        "train_extra": ["thai-ipa/train.jsonl", "thai-ipa/augmented_epitran.jsonl"],
+        "val": "thai-ipa-expanded/val.jsonl",
+        "test": "thai-ipa-expanded/test.jsonl",
+        "eval_test": "thai-ipa/test.jsonl",
+        "out": "secryst_thai_g2p_distill_small/run-003",
+        "mode": "sequence",
+        "note": "eval-only spec for tha-g2p-small-1.0 (client tier)",
+    },
     "ara-diac-small": {
         # r5 paragraph-context teacher (2.68 DER-CE windowed @1400B,
         # RELEASE-FROZEN) -> ByT5-small student. Contract decode is

src/gpu/modal_export.py

@@ -113,6 +113,15 @@
         "test_data": "thai-ipa/test.jsonl",
         "probe": "สวัสดี",
     },
+    "tha-g2p-small": {
+        "volume": "/volumes/secryst-checkpoints",
+        "checkpoint": "secryst_thai_g2p_distill_small/run-003/best",
+        "metadata": "models/tha-g2p-small/tha-g2p-small-1.0.metadata.yaml",
+        "readme": "models/tha-g2p-small/tha-g2p-small-1.0.README.md",
+        "test_volume": "/datasets/secryst",
+        "test_data": "thai-ipa/test.jsonl",
+        "probe": "สวัสดี",
+    },
     "fas-g2p": {
         "volume": "/volumes/persian-checkpoints",
         "checkpoint": "persian_g2p/run-001/best",