Skip to content

khm-latn-1.0 — first interscript-ml model release

Choose a tag to compare

@ronaldtse ronaldtse released this 16 Aug 10:24
· 758 commits to main since this release

khm-latn-1.0 — first interscript-ml release

Khmer → Latin transliteration. ByT5-small, byte-level (canonical ByT5
table), IMF v1 zip with KV-cache decoder (plain fallback included).
Trained from secryst train_khmer_byt5.py run-001
(secryst-checkpoints:/khmer_byt5/run-001/best).

field value
task translit (Khmr → Latn)
artifact khm-latn-1.0-fp32.zip (1.32 GiB)
zip sha256 55993d473a2ed9489058779cad7e115db05e4020085b71616468cfae4f2f65cb
CER / EM 27.42 / 59.66 — greedy, 895 held-out pairs (secryst/docs/RESULTS.md#khmer-transliteration-2026-08-14)
parity cer_delta 0.0pp on 895 samples (fp32, token-exact vs torch reference)
license BSD-3-Clause

Precision variants fp16/int8 are exported and on the pipeline volume but
NOT in this release: they measure 0.43pp / 0.84pp CER delta against the
0.2pp release bar (quantization noise under greedy decoding on a weak
model); threshold policy for lossy precisions pending.

Fetch (any of the three runtimes):

from interscript_ml import Model
model = Model.load("khm-latn-1.0")
model.translate("ភាសា")  # -> "pheasaea"

Resolution contract: models.yaml at the repo root (raw GitHub / jsDelivr
mirrors) carries the channel URL + whole-file sha256; runtimes download,
verify, and cache. Ruby (secryst gem) and TypeScript (@interscript/ml)
bindings implement the same contract.