Skip to content

docs: WO68 — expansion B negative (mis-design), B2 corrected #397

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 — expansion B negative (mis-design), B2 corrected #397

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
27 changes: 27 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 @@ -118,3 +118,30 @@ retrains; the Khmer regime (tiny gold, huge unlabeled) is the case weak
supervision exists for, with held-out validation mandatory against the WO64
correlated-teacher negative. This path makes CER < 15 reachable without
SEAlang / Chuon Nath.


### Expansion attempt 2 — stage A done, stage B negative (mis-design), B2 in flight (2026-10-30)

SBBIC 75,039-word list labeled with the shipped khm-phonemic-1.0 teacher
(house runtime, checkpointed rotating parts): **73,954 labels** persisted
(bucket khmer-expansion/sbbic-labeled.jsonl, 4.56 MB; 10,046 resume dupes
dropped; 1,038 unique words dropped by the labeler's empty-output guard —
words where the teacher emits nothing; 1.4%, acceptable).

**Stage B NEGATIVE — an artifact of target-convention mismatch** (my
script bug, recorded so it is never re-read as evidence): it trained on
the teacher's Latin romanizations (SBBIC weak labels) but evaluated
against the Chuon Nath IPA holdout. CER 0.735 / EM 0 vs the 0.245
reference measures the Latin-vs-IPA convention gap, not expansion value.
NO_BETTER_THAN_TEACHER from that run carries no information.

**Stage B2 (corrected, running as run-082d-khm-expansion2)**: matched
surface — SBBIC weak Latin labels + the translit corpus train anchor
(exact run-081 protocol: sorted pairs, seed-42 shuffle, 5/5/90), with
translit words dropped from the weak set (contamination control). The
teacher is scored on the IDENTICAL 895-row translit test split in-job
(house runtime on the shipped artifact), so the verdict is
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.
Loading