Skip to content
Merged
Changes from all commits
Commits
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
21 changes: 21 additions & 0 deletions TODO.sota/68-b7-khmer-lexicon.md
Original file line number Diff line number Diff line change
Expand Up @@ -63,3 +63,24 @@ corpora/khm-latn.yaml provenance, and the retrain is unblocked.
raw.githubusercontent fetch), port the Modal training script to HF Jobs
(--detach, >=4h, step ckpts), retrain ByT5, eval against khm-latn-1.0's
27.4 CER on the same protocol, ship via the strict publish path if better.


## Lineage recovery VERIFIED by reproduction (2026-10-29)

run-081-khm-latn-recovery (HF job rhtse/6ae32d5e3fb22a235dd5896f, seed 42,
legacy recipe; first attempt CANCELED externally at epoch 4, attempt 2 of 3
completed): held-out 5% test split, 895 rows —

- greedy: EM 58.21%, CER 28.13%
- beam4: EM 58.44%, **CER 27.49%**

Against khm-latn-1.0's 27.42 CER this is reproduction within noise: the
secryst/data-khmer-translit corpus is confirmed as the khm-latn training
lineage, and the provenance pin in corpora/khm-latn.yaml is now verified by
reproduction (the strongest evidence class). Best checkpoint:
bucket Interscript/isx-training/khmer-translit/run-081-recovery/best.

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.
Loading