Skip to content

docs: WO64 — B3 is a re-train campaign #395

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 30, 2026
Merged

docs: WO64 — B3 is a re-train campaign #395

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
20 changes: 20 additions & 0 deletions TODO.sota/64-b3-arabic-sota.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -101,3 +101,23 @@ news specialist (ara-diac-news-1.0)}; init = the shipped r7 artifact. The
re-baselines (first-alt WER is measured against the r7 init's own
run), but every input is pinned and checksummed today. Recommend B, with
A opportunistically revisited if the bucket entry reappears.


### Storage verdict (2026-10-30): B3 is a re-train campaign

A full recursive dump of rhtse/jobs-artifacts (2,257 entries) contains ZERO
run-024 entries — the run-024-arabic-r8-* checkpoint dirs that appeared in
several listing greps earlier the same day have since vanished, as did the
p047 export. The bucket layer is lossy for the WO35/r8-era artifacts; no
launch can be built on restaging them.

**Definitive path (every input stable):** re-train the arms from the
recorded recipes against the staged corpus — `train_arabic_r7.py` is present
and the corpus is mounted from the Interscript/arabic-r8 dataset repo
(owner-approved license review 2026-10-18; also hosting SadeedDiac-25 and
the WikiNews-2024 multiref eval splits). Additionally documented corpus
options: qcri/advancing-arabic-diacritization (EMNLP 2025, CC BY-NC-SA —
refined datasets + the SadeedDiac-25 benchmark) and Misraj/Sadeed_Tashkeela
(gated, ~doubles the corpus). Cost: r7 re-train + 2-arm relabel +
continue-train ~= 2-3 GPU-days. The re-baselined gate (first-alt WER against
the re-trained init) is recorded before launch per the Option-B consequence.
Loading