Skip to content

docs: khm lineage recovery + persian source dossier #368

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 25, 2026
Merged

docs: khm lineage recovery + persian source dossier #368

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
12 changes: 12 additions & 0 deletions TODO.sota/66-b5-persian-registers.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,3 +7,15 @@ Method: register-stratified eval sets (build from poetry corpora,
colloquial sources); probe the scar (docs 0.52% lesson: corpus-level vs
row-mean declared); targeted data if registers degrade; VRR.
Acceptance: register-stratified external-protocol numbers published.


## Source dossier (2026-10-25, for owner approval)

| source | register | license | note |
|---|---|---|---|
| Ganjoor (ganjoor.net) | poetry (Hafez/Saadi/…) | CC BY-SA 3.0 (site-stated) | the largest digitized divan corpus; attribution + share-alike |
| fa-Wikisource | poetry + classical prose | CC BY-SA 4.0 | attributable, steady quality |
| OSCAR-fa | colloquial/web | Common-Crawl-derived (CC0-shaded tooling) | register = modern web; quality filters needed |

Owner picks the sources -> corpora/<x>.yaml with sha256 pins ->
the register-hold arm (target: CER < 0.7 across registers).
11 changes: 11 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 @@ -26,3 +26,14 @@ prefix conditioning); fp16 zip hygiene (khm fp32 corruption lesson);
conformance fixtures for the new artifact.
Acceptance: gates green; hit-rate + CER on held-out; release via the
strict publish path.


## Lineage recovery (2026-10-25)

khm-latn-1.0 traces to the **secryst project**
(`~/src/interscript/secryst`; README documents
`scripts/export_onnx_byt5.py <hf_checkpoint_dir> khm-latn-byt5.zip`).
The training corpus is re-derivable from the secryst repo's data
pipeline + khmer-dict-spice — the recovery question is now concrete:
pin the exact secryst-era corpus files, run them through
corpora/khm-latn.yaml provenance, and the retrain is unblocked.
Loading