Skip to content

docs: WO68 — B2 EXPANSION_PAYS; attempt 3 latn-convention design #398

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: WO68 — B2 EXPANSION_PAYS; attempt 3 latn-convention design #398

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
21 changes: 21 additions & 0 deletions TODO.sota/68-b7-khmer-lexicon.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -145,3 +145,24 @@ student-vs-teacher on the same rows and surface. Gate: student CER <
teacher CER on that split -> EXPANSION_PAYS. Note the translit test's
anchor CER band is ~27.4 (run-081 reproduction); the teacher's own CER
on the same rows is the number that matters.


**Stage B2 VERDICT: EXPANSION_PAYS** (2026-10-30, run-082d). Matched-surface
result on the translit romanization test split (895 rows, exact run-081
protocol): student CER 40.25% / EM 52.4% vs the phonemic teacher
(khm-phonemic-1.0, same rows in-job) 47.15% / EM 44.6% — the 57K SBBIC
weak labels + 16K translit anchor beat the teacher by 6.9 CER points on
its own surface. Caveat carried forward: this compares against the
PHONEMIC teacher; the khm-latn lineage's own band is 27.4, and the
translit-surface score of a convention-blended model is not a lineage
benchmark.

Attempt 3 (running): the convention-correct experiment — SBBIC labeled
with the SHIPPED khm-latn-1.0 artifact (run-082e-khm-latn-label, 24
CPU processes on a100-large, rotating parts, sha256-verified Release
fetch; plain decoders are batch=1-traced but words are tiny) -> weak
labels in the translit convention -> train byt5-small (weak + translit
train anchor, exact run-081 split) -> gate: student CER < teacher CER
on the identical 895 test rows (the lineage comparison vs 27.4).
Verdict script: train_khm_expansion3.py (stage-b3 outputs under
bucket khmer-latn/).
Loading