Skip to content

docs: WO65 — Urdu oracle relabel staged for post-B3 launch #402

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ronaldtse merged 1 commit into from
Oct 31, 2026
Merged

docs: WO65 — Urdu oracle relabel staged for post-B3 launch #402

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 22 additions & 0 deletions TODO.sota/65-b4-urdu-sota.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,3 +29,25 @@ Engineering, UET Lahore). Owner-side steps enumerated for the record:
5. On license-green: retrain urd-g2p on the licensed corpus, re-run WO51
gates, flip card/index from research-only, ship. Separately: whether the
CLE corpus also feeds the urd-diac 3.57 -> < 3 push is its own WO.


### WO65 stage-1 staged (2026-10-31): launch command ready

The B3 oracle machinery generalizes to Urdu (oracle_relabel_onnx.py:
D_ARM1_* env primary arm + D_ROWS_JSONL pre-paired corpus mode, PR'd in
interscript-train #141). Launch after the B3 arms free the a100s:

- Arms (both SHIPPED, stable): urd-diac-1.0 (urd-diac-1.0-fp32.zip,
release tag urd-diac-1.0, 1.42GB fp32) + urd-diac-base-1.0
(urd-diac-base-1.0-fp16.zip, tag urd-diac-base-1.0, 1.35GB fp16,
different lineage = decorrelated). Stage both zips to bucket
urd-arms/ first (sha256 from models.yaml).
- Corpus: bucket rhtse/jobs-artifacts/rababa-urdu-cc174b96 revision
data-diacrit -> mount /data-dia; rows = data-diacrit/train.jsonl
(573,130 rows, sha256 in corpora/urd-dia.yaml, owner-approved
license review 2026-10-18). D_ROWS_JSONL=/data-dia/data-diacrit/
train.jsonl D_SUBSET=200000 D_MAXB=300 D_BS=48.
- Run dir: D_RUN=run-088-urd-oracle; two parallel D_ARM_ONLY jobs as
in B3. Then stage 2 = train_oracle_student_hf.py adapted (O_INIT
google/byt5-base; corpus mounts change) -> gate CER < 3.57 baseline,
target CER < 3 (WO65 method: decorrelated teacher arms only).
Loading