Skip to content

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

Merged
ronaldtse merged 1 commit into
mainfrom
docs/wo64-retrain
Oct 30, 2026
Merged

ronaldtse merged 1 commit into
mainfrom
docs/wo64-retrain

Conversation

@ronaldtse

Copy link
Copy Markdown
Contributor

Summary

Final WO64 verdict: the bucket layer is lossy for the WO35/r8-era artifacts (run-024 checkpoint dirs surfaced in listings and vanished within the day; the full 2,257-entry dump contains zero run-024 entries; the p047 export appeared once and vanished). B3 is therefore a re-train campaign on stable inputs only: the recorded recipes (train_arabic_r7.py present), the corpus 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), and the documented additional corpora (EMNLP 2025 QCRI refined set, CC BY-NC-SA; Misraj Sadeed_Tashkeela, gated). Cost estimate: 2-3 GPU-days; the re-baselined gate (first-alt WER against the re-trained init) is recorded before launch.

@ronaldtse
ronaldtse merged commit 4661765 into main Oct 30, 2026
8 of 12 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