urd-g2p-base-1.0

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 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.

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.