release: fas-g2p-1.0 (fp32, parity cer_delta 0.0pp on 1000 samples)

Author: Ronald Tse

models.yaml

@@ -117,4 +117,22 @@ models:
       - {name: der_teacher_greedy, value: 24.79, source: interscript/interscript-ml docs/RESULTS.md#heb-diac-small-1.0}
       - {name: der_student_greedy, value: 30.37, source: interscript/interscript-ml docs/RESULTS.md#heb-diac-small-1.0}
     parity: {samples: 1864, cer_delta: 0.0}
+  fas-g2p-1.0:
+    task: g2p
+    scripts: [Arab, Latn]
+    precision: fp32
+    filename: fas-g2p-1.0-fp32.zip
+    parts:
+      - url: https://github.com/interscript/interscript-ml/releases/download/fas-g2p-1.0/fas-g2p-1.0-fp32.zip.part-00
+        sha256: 6c475e53f3c7ef3edcef5a24d2d81f51d13ddf4c417de98440943c9a32b03e55
+        size: 1500000000
+      - url: https://github.com/interscript/interscript-ml/releases/download/fas-g2p-1.0/fas-g2p-1.0-fp32.zip.part-01
+        sha256: c5656c649412f41e201c567e002514d31390b636d4650d9dcb6dfed1973fb3d5
+        size: 1280254270
+    sha256: b53fd69196127efcc596935d24e135b3c897006872c1646c674e70f3c0d9a0e3
+    size: 2780254270
+    metrics:
+      - {name: cer, value: 1.6, source: interscript/interscript-ml docs/RESULTS.md#fas-g2p-1.0}
+      - {name: sb_homograph_ezafe_norm, value: 77.34, source: interscript/interscript-ml docs/RESULTS.md#fas-g2p-1.0}
+    parity: {samples: 1000, cer_delta: 0.0}
     license: BSD-3-Clause

models/fas-g2p/fas-g2p-1.0.metadata.yaml

@@ -8,22 +8,24 @@ opset: 14
 decoder: kv
 precision: fp32
 license: BSD-3-Clause
-trained_from: >-
-  persian_g2p v1 (ByT5-small, byte-level, RELEASE-FROZEN); checkpoint
-  persian-g2p-checkpoints:/checkpoints/persian_g2p/run-001/best.
-  Shipped directly as the client-tier model — the v1 teacher is already
-  byte-level ByT5-small (no distillation needed; GRPO/RAFT variants and
-  the v5/mapped representation line are closed negative).
+trained_from: persian_g2p v1 (ByT5-small, byte-level, RELEASE-FROZEN); checkpoint
+  persian-g2p-checkpoints:/checkpoints/persian_g2p/run-001/best. Shipped directly
+  as the client-tier model — the v1 teacher is already byte-level ByT5-small (no distillation
+  needed; GRPO/RAFT variants and the v5/mapped representation line are closed negative).
 metrics:
-  - name: cer
-    value: 1.6
-    protocol: >-
-      greedy decode, editdistance CER on the v1 test split
-      (persian-g2p-datasets:/persian-g2p/test.jsonl)
-    source: interscript/interscript-ml docs/RESULTS.md#fas-g2p-1.0
-  - name: sb_homograph_ezafe_norm
-    value: 77.34
-    protocol: >-
-      SentenceBench homograph accuracy, ezafe-normalized
-      (eval_sentencebench.py protocol; published Homo-GE2PE: 76.89%)
-    source: interscript/interscript-ml docs/RESULTS.md#fas-g2p-1.0
+- name: cer
+  value: 1.6
+  protocol: greedy decode, editdistance CER on the v1 test split (persian-g2p-datasets:/persian-g2p/test.jsonl)
+  source: interscript/interscript-ml docs/RESULTS.md#fas-g2p-1.0
+- name: sb_homograph_ezafe_norm
+  value: 77.34
+  protocol: 'SentenceBench homograph accuracy, ezafe-normalized (eval_sentencebench.py
+    protocol; published Homo-GE2PE: 76.89%)'
+  source: interscript/interscript-ml docs/RESULTS.md#fas-g2p-1.0
+parity:
+  samples: 1000
+  cer_delta: 0.0
+sha256:
+  decoder-kv.onnx: ca5845e41ee7b0be050801a136399a0f1fd78f66973d68a68bfc09c97fd1550c
+  decoder.onnx: 0301054d602736bb0333c3c10c2836a74764c28a61e49b8472e7c637defd7f31
+  encoder.onnx: 811c0fd2e3e275dc00b49d299d4c3091cc841e4b87c775a528a0fa88fd36f488