diff --git a/TODO.sota/64-b3-arabic-sota.md b/TODO.sota/64-b3-arabic-sota.md index c0b39ab..61bfde9 100644 --- a/TODO.sota/64-b3-arabic-sota.md +++ b/TODO.sota/64-b3-arabic-sota.md @@ -68,3 +68,19 @@ owner-visible, with two options: 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). + + +### Option A scoping continued (2026-10-30) + +With the owner unlock, Option A is the chosen path (faithful gate lineage). +Recipe audit: WO39 records run-047's CORPUS recipe exactly (MIX=qcri, +R8_GOLD=none, PSEUDO_CAP=900K, seed-7 article shuffle, WINDOW=600, +STRIP_ALT targets) — stage 1 is runnable once the arm checkpoints restage. +The large-strip TRAINING script itself is not in interscript-train (the +r-series scripts cover byt5/r3/r5/r6/r7/r8/r9/plane/raft); identifying or +reconstructing it is the remaining archaeology. The r7 lineage is fully +reproducible today (train_arabic_r7.py present). Execution order: (1) +restage p029 from the shipped ara-diac-news release; (2) re-derive +run-047-large-strip from the recovered recipe (the stage-2 init); (3) +re-derive or substitute p040; (4) run the WO39 chain. Steps 1-2 start next +session as the focused task. diff --git a/TODO.sota/68-b7-khmer-lexicon.md b/TODO.sota/68-b7-khmer-lexicon.md index 9713866..1aa95c0 100644 --- a/TODO.sota/68-b7-khmer-lexicon.md +++ b/TODO.sota/68-b7-khmer-lexicon.md @@ -105,3 +105,16 @@ Alternative license-clear expansion paths identified, attempt 1 launched: - FLEURS-Khmer alignment (heavier, parked). The expansion path makes CER < 15 reachable WITHOUT SEAlang/Chuon Nath. + + +## Attempt 2 launched: self-training on the license-clear SBBIC list (2026-10-29) + +khm-phonemic-1.0 (our shipped 24.5-CER model — not an LLM; the LLM-teacher +law is not implicated) labels the 75,039-word SBBIC list (LGPL-2.1, in +khmer-dict-spice, staged to the training bucket) with IPA readings plus mean +token log-prob confidence (run-082b-khm-sbbic-label, HF job +rhtse/6ae33c0441e90535df78b820). Stage B filters hard on confidence and +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.