Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# urd-g2p-base-1.0
|
| 2 |
+
|
| 3 |
+
Urdu grapheme-to-phoneme, base tier (run-059): byt5-base seq2seq on
|
| 4 |
+
the full 610K-pair humair025 corpus. IMF v1 fp32 (2.16 GB).
|
| 5 |
+
|
| 6 |
+
## Metrics (full 12,699-row test split, raw char CER, row-mean)
|
| 7 |
+
|
| 8 |
+
CER **15.65** (urd-g2p-1.0 small tier: 15.78 baseline) — gate PASSED.
|
| 9 |
+
Normalized CER 10.75.
|
| 10 |
+
|
| 11 |
+
## Artifact parity (WO51 contract)
|
| 12 |
+
|
| 13 |
+
Single-row greedy decode, the runtime's actual decode mode: 500-row
|
| 14 |
+
probe recorded in the artifact metadata (the pre-WO51 batched
|
| 15 |
+
harness numbers were invalid — batched decode without an attention
|
| 16 |
+
mask measures nothing).
|
| 17 |
+
|
| 18 |
+
## License — RESEARCH-ONLY
|
| 19 |
+
|
| 20 |
+
The training corpus (humair025/urdu-g2p-dictionary) is
|
| 21 |
+
**non-commercial** (verified from the dataset card,
|
| 22 |
+
github.com/humair-m/urdu-g2p LICENSE). These weights are therefore
|
| 23 |
+
released **for research use only** — owner decision 2026-10-18
|
| 24 |
+
(recorded in `corpora/urd-g2p-dict.yaml` license_review). Code:
|
| 25 |
+
BSD-3-Clause.
|