Release: urd-g2p-base-1.0 (research-only)

cards/urd-g2p-base-1.0.md

@@ -3,8 +3,23 @@
 Urdu grapheme-to-phoneme, base tier (run-059): byt5-base seq2seq on
 the full 610K-pair humair025 corpus. IMF v1 fp32 (2.16 GB).
 
-## Metrics (full 11,940-row test split, raw char CER)
+## Metrics (full 12,699-row test split, raw char CER, row-mean)
 
 CER **15.65** (urd-g2p-1.0 small tier: 15.78 baseline) — gate PASSED.
-Normalized CER 10.75. Shipped-bytes parity: ONNX fp32 vs torch
-cer_delta 0.024 on the 300-row sample, 70% row-exact.
+Normalized CER 10.75.
+
+## Artifact parity (WO51 contract)
+
+Single-row greedy decode, the runtime's actual decode mode: 500-row
+probe recorded in the artifact metadata (the pre-WO51 batched
+harness numbers were invalid — batched decode without an attention
+mask measures nothing).
+
+## License — RESEARCH-ONLY
+
+The training corpus (humair025/urdu-g2p-dictionary) is
+**non-commercial** (verified from the dataset card,
+github.com/humair-m/urdu-g2p LICENSE). These weights are therefore
+released **for research use only** — owner decision 2026-10-18
+(recorded in `corpora/urd-g2p-dict.yaml` license_review). Code:
+BSD-3-Clause.

docs/RESULTS.md

@@ -2357,3 +2357,12 @@ parity under the WO51 single-row contract: ONNX fp16 CER 3.64,
 500/500 rows identical to torch, three-way bf16/fp32/ORT agreement
 1.0. Shipped runtime reproduces 12/12 frozen conformance vectors
 byte-exactly. Data: 573,130 pairs (urd-dia registry, sha256-pinned).
+
+## run-059 urd-g2p base gate pass
+
+Gate: full 12,699-row test split, raw char CER, row-mean — **15.65**
+(beat urd-g2p-1.0 small tier 15.78); normalized CER 10.75. Data:
+610K pairs (humair025/urdu-g2p-dictionary via rababa-urdu/data,
+sha256-pinned in the urd-g2p-dict registry). Training data is
+non-commercial: weights ship RESEARCH-ONLY (owner decision
+2026-10-18).

models.yaml

@@ -451,15 +451,22 @@ models:
     task: g2p
     scripts: [Arab, IPA]
     precision: fp32
-    filename: urd-g2p-base-1.0.zip
-    url: https://github.com/interscript/interscript-models/releases/download/urd-g2p-base-1.0/urd-g2p-base-1.0.zip
-    sha256: 14c2ebf81999f24b624c7f3c9d223a16c36fee465863785b6ceea88b3d1f371a
-    size: 2159598592
-    metrics:
-      - {name: cer, value: 15.65, protocol: "greedy decode, full 11940-row test split (WO37 protocol)", source: interscript/interscript-models docs/RESULTS.md#run-059-urd-g2p-base-gate-pass}
-      - {name: cer_normalized, value: 10.75, protocol: "greedy decode, full 11940-row test split (WO37 protocol)", source: interscript/interscript-models docs/RESULTS.md#run-059-urd-g2p-base-gate-pass}
-    parity: {samples: 300, cer_delta: 0.024}
+    filename: urd-g2p-base-1.0-fp32.zip
+    parts:
+      - url: https://github.com/interscript/interscript-models/releases/download/urd-g2p-base-1.0/urd-g2p-base-1.0-fp32.zip.part-00
+        sha256: e2ffe9bf8114e26b731d4469a262b2d17d0a8f6ceced611c641e7d69bca804de
+        size: 1500000000
+      - url: https://github.com/interscript/interscript-models/releases/download/urd-g2p-base-1.0/urd-g2p-base-1.0-fp32.zip.part-01
+        sha256: 72695f23f42f8eef9d81c8284d02cbf743986fd836aa1568cf11a08fc75acd36
+        size: 659598918
+    sha256: d3e35dd435bf18fedec03d4a15bba097fca743737b729312fa08f4fdcac24282
+    size: 2159598918
+    metrics:
+      - {name: cer, value: 15.65, protocol: "greedy decode, full 12699-row test split, row-mean", source: interscript/interscript-models docs/RESULTS.md#run-059-urd-g2p-base-gate-pass}
+      - {name: cer_normalized, value: 10.75, protocol: "greedy decode, full 12699-row test split, row-mean", source: interscript/interscript-models docs/RESULTS.md#run-059-urd-g2p-base-gate-pass}
+    parity: {samples: 500, cer_delta: 0.0}
     license: BSD-3-Clause
+    license_note: RESEARCH-ONLY — non-commercial training data (owner-approved 2026-10-18; see card)
   ara-diac-news-1.1:
     task: diacritization
     scripts: [Arab, Arab]

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

@@ -0,0 +1,28 @@
+format: imf-v1
+id: urd-g2p-base-1.0
+task: g2p
+source_script: Arab
+target: IPA
+tokenizer: bytes
+opset: 14
+decoder: plain
+precision: fp32
+license: BSD-3-Clause
+trained_from: byt5-base, 610K humair025 pairs, run-059 (TODO.sota WO38)
+metrics:
+- name: cer
+  value: 15.65
+  source: interscript/interscript-models docs/RESULTS.md#run-059-urd-g2p-base-gate-pass
+  protocol: greedy decode, full 12699-row test split, row-mean CER (WO51 single-row
+    artifact contract)
+- name: cer_normalized
+  value: 10.75
+  source: interscript/interscript-models docs/RESULTS.md#run-059-urd-g2p-base-gate-pass
+  protocol: greedy decode, full 12699-row test split, row-mean CER (WO51 single-row
+    artifact contract)
+sha256:
+  encoder.onnx: 34263b20c4aef190ec8b5b0a68be04f684147a7c2e954e9f89f618525dd835bb
+  decoder.onnx: 0a21fdbb9783e2aed6d75c63370036d9bb2665d26404ed627bf5883b2c39e953
+parity:
+  samples: 500
+  cer_delta: 0.0