From c84b71d5d6a300c11062888bcbf19ba89fe44ce2 Mon Sep 17 00:00:00 2001
From: Ronald Tse <ronald.tse@gmail.com>
Date: Thu, 3 Sep 2026 15:29:20 +0200
Subject: [PATCH] =?UTF-8?q?feat(export):=20ara-diac-layerdrop-1.0=20?=
 =?UTF-8?q?=E2=80=94=20the=20int4=20browser-budget=20rung?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

~190M depth-cut student (verbatim layer copies), full-set 5.78
[3.03-3.49]. The int4 zip (~95MB) is the first browser-budget
diacritization artifact; margin-gated before ship. (TODO.publish-client 02)
---
 .../ara-diac-layerdrop-1.0.README.md          | 16 ++++++++++
 .../ara-diac-layerdrop-1.0.metadata.yaml      | 30 +++++++++++++++++++
 src/gpu/modal_export.py                       |  9 ++++++
 3 files changed, 55 insertions(+)
 create mode 100644 models/ara-diac-layerdrop/ara-diac-layerdrop-1.0.README.md
 create mode 100644 models/ara-diac-layerdrop/ara-diac-layerdrop-1.0.metadata.yaml

diff --git a/models/ara-diac-layerdrop/ara-diac-layerdrop-1.0.README.md b/models/ara-diac-layerdrop/ara-diac-layerdrop-1.0.README.md
new file mode 100644
index 0000000..c791871
--- /dev/null
+++ b/models/ara-diac-layerdrop/ara-diac-layerdrop-1.0.README.md
@@ -0,0 +1,16 @@
+# ara-diac-layerdrop-1.0
+
+The depth-cut client rung: ByT5-small with the encoder halved
+(layers copied verbatim — width survives, depth is the compressible
+axis). Muon + r7 labels + 6 epochs, the full lever set.
+
+- Full-set DER: **5.78** (delta CI [3.03, 3.49]; full-depth peer 4.57)
+- ~190M parameters; int8 ~190MB; **int4 ~95MB — the browser budget**
+- The capacity law in one artifact: halving depth costs 1.21pp where
+  every width surgery collapsed
+
+```python
+from interscript_ml import Model
+model = Model.load("ara-diac-layerdrop-1.0")
+model.translate("قوله فحكمها في الوفاة")
+```
diff --git a/models/ara-diac-layerdrop/ara-diac-layerdrop-1.0.metadata.yaml b/models/ara-diac-layerdrop/ara-diac-layerdrop-1.0.metadata.yaml
new file mode 100644
index 0000000..26cc5b6
--- /dev/null
+++ b/models/ara-diac-layerdrop/ara-diac-layerdrop-1.0.metadata.yaml
@@ -0,0 +1,30 @@
+format: imf-v1
+id: ara-diac-layerdrop-1.0
+task: diacritization
+source_script: Arab
+target: Arab
+tokenizer: bytes
+opset: 14
+decoder: kv
+precision: fp32
+license: BSD-3-Clause
+trained_from: 'the depth-cut rung: ByT5-small with the encoder halved
+  12->6 (surviving layers copied VERBATIM from pretraining - both width-cut
+  approaches collapsed at 74.68/82.96 while this works), Muon optimizer, r7
+  teacher labels (the 2.0 recipe), 6 epochs. ~190M parameters (63% of
+  ByT5-small). Full-set windowed DER-CE 5.784, delta CI [3.033, 3.491] vs
+  in-run teacher 2.2921 — the depth premium over the full-depth G2a peer
+  (4.5701, CI [1.911, 2.352]) is 1.21pp with non-overlapping intervals.
+  The int4 variant (~95MB) is the browser-budget tier; margin-gated before
+  ship. Checkpoint
+  rababa-checkpoints:/rababa_arabic_distill_small/run-009-layerdrop-6ep/best.'
+metrics:
+- name: der_teacher_fullset
+  value: 2.2921
+  protocol: windowed DER-CE (1400-byte windows, word-boundary split, greedy,
+    haraqat-projected, Misraj evaluator); full 1,200-paragraph SadeedDiac-25
+  source: interscript/interscript-ml docs/RESULTS.md#layerdrop-6ep
+- name: der_student_fullset
+  value: 5.784
+  protocol: same harness; paired bootstrap delta 3.2455 [3.033, 3.491]
+  source: interscript/interscript-ml docs/RESULTS.md#layerdrop-6ep
diff --git a/src/gpu/modal_export.py b/src/gpu/modal_export.py
index 1836db1..cbe368a 100644
--- a/src/gpu/modal_export.py
+++ b/src/gpu/modal_export.py
@@ -132,6 +132,15 @@
         "test_data": "arabic-sadeed-imf/test.jsonl",
         "probe": "قوله",
     },
+    "ara-diac-layerdrop": {
+        "volume": "/volumes/rababa-checkpoints",
+        "checkpoint": "rababa_arabic_distill_small/run-009-layerdrop-6ep/best",
+        "metadata": "models/ara-diac-layerdrop/ara-diac-layerdrop-1.0.metadata.yaml",
+        "readme": "models/ara-diac-layerdrop/ara-diac-layerdrop-1.0.README.md",
+        "test_volume": "/datasets/rababa",
+        "test_data": "arabic-sadeed-imf/test.jsonl",
+        "probe": "قوله",
+    },
     "urd-diac": {
         "volume": "/volumes/urdu-diacrit-checkpoints",
         "checkpoint": "urdu_diacrit/run-001/best",