Skip to content

docs: WO64 — oracle-distill launch scoping - #389

Merged
ronaldtse merged 2 commits into
mainfrom
docs/wo64-oracle-launch-scope
Oct 29, 2026
Merged

ronaldtse merged 2 commits into
mainfrom
docs/wo64-oracle-launch-scope

Conversation

@ronaldtse

Copy link
Copy Markdown
Contributor

Summary

Records in WO64 that the WO39 oracle-distill chain is script-complete (stage-1 relabel, stage-2 continue-train with SadeedDiac + WikiNews evals; gate first-alt WER < 3.5607, stretch <= 2.70) and that the sole launch prerequisite is staging the three WO35 arm checkpoints plus the QCRI train_data mount onto HF storage. The legacy Modal volumes list only thai/rababa-era artifacts; the 78 timestamped job-sync dirs on rhtse/jobs-artifacts are the candidate location; if the checkpoints are unrecoverable, two of three arms restage from shipped releases (ara-diac-news-1.0, the r7 lineage) and the stage-2 init question becomes an owner-visible design decision.

Ronald Tse added 2 commits October 29, 2026 19:18
…t identified and staged

The WO68 lineage recovery: the training corpus is exactly
secryst/data-khmer-translit (17,911 pairs), staged to the training
bucket with sha256, retiring the raw.githubusercontent fetch. The
upstream repo has no license file, so the ship gate stays owner-held.
Recovery run-081 launched (seed 42, legacy recipe).
@ronaldtse
ronaldtse merged commit 85bcaaf into main Oct 29, 2026
7 of 11 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant