urd-diac-base-1.0

Urdu diacritization, base tier (run-066): byt5-base seq2seq on the full 573,130-pair urdu-diacrit corpus. IMF v1 fp16.

Metrics (full 11,940-row test split, raw char CER, row-mean)

CER 3.57 (urd-diac-1.0 small tier: 3.77 baseline) — gate PASSED. Word accuracy 69.8%.

Artifact parity (WO51 contract)

Single-row greedy decode, the runtime's actual decode mode: ONNX artifact CER 3.64 — identical to torch on 500/500 rows (agreement 1.0; three-way bf16/fp32/ORT agreement 1.0). Batched decode without an attention mask is outside the artifact contract (CONFORMANCE.md).

Conformance

The shipped runtime reproduces all 12 frozen vectors byte-exactly (conformance/urd-diac-base-vectors.json).

Data / license

Weights depend on the original training-data license (the urdu-diacrit corpus; provenance in corpora/PROVENANCE.md, owner-approved ship decision 2026-10-18). Code and weights: BSD-3-Clause.