Skip to content

docs: WO68 expansion attempt 1 launched; WO64 archaeology concluded #391

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 29, 2026
Merged

docs: WO68 expansion attempt 1 launched; WO64 archaeology concluded #391

Show file tree
Hide file tree
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
17 changes: 17 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 @@ -51,3 +51,20 @@ Findings of the audit:
short re-derivation from its recorded recipe, or the stage-2 init moves
to the closest recoverable ancestor — an owner-visible design decision
recorded before any run.


### Archaeology CONCLUDED (2026-10-29): arms unrecoverable from buckets

The 78 timestamped job-sync dirs on rhtse/jobs-artifacts contain only script
copies (train_arabic_r8_hf.py et al.) — no checkpoints. The WO35 arm
checkpoints are unrecoverable from our storage. The launch decision is now
owner-visible, with two options:

- **Option A (restage + re-derive):** restage the two shipped arms
(ara-diac-news-1.0/1.1, the r7 canonical lineage) from pinned releases;
re-derive run-047-large-strip/best from its recorded recipe (a short
continue-train of its own) — then run the WO39 chain as designed.
- **Option B (init substitution):** stage-2 continue-trains from the closest
recoverable ancestor instead of run-047/best — cheaper but changes the
gate lineage (the 3.5607 gate was measured ON run-047/best; a different
init re-baselines the comparison).
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 @@ -84,3 +84,24 @@ Consequence for the target: recovery is closed, but reproduction at ~27.5
also confirms the corpus ceiling — **CER < 15 requires data expansion**, so
the owner's licensing decision (SEAlang / full Chuon Nath IPA, or a blessed
corpus-release license opening other sources) is now THE B7 blocker.


## Data expansion attempt 1: Wiktionary/kaikki (2026-10-29, launched)

The owner reports SEAlang / full Chuon Nath IPA are not findable/obtainable.
Alternative license-clear expansion paths identified, attempt 1 launched:

- **kaikki/Wiktionary Khmer** (CC BY-SA 4.0, the tha-lexicon-kaikki-1.0
precedent): the dump exists at
kaikki.org/dictionary/Khmer/kaikki.org-dictionary-Khmer.jsonl. Job
run-082-khm-wikt-expansion extracts word->IPA and word->romanization
pairs for Khmer-script headwords.
- **Self-training expansion** (attempt 2, designed): khm-phonemic-1.0
(24.5 CER) labels the 72,048 SBBIC words (license-clear, in
khmer-dict-spice) -> confidence-filtered gold expansion. Caution: the
noisy-student negative (WO64) was correlated-teacher Arabic — the Khmer
regime (tiny gold, huge unlabeled) is the case self-training is for;
confidence filtering + held-out VRR-style validation mandatory.
- FLEURS-Khmer alignment (heavier, parked).

The expansion path makes CER < 15 reachable WITHOUT SEAlang/Chuon Nath.
Loading