HPLT v2.0 Ukrainian→English — CTranslate2 int8
HPLT/translate-uk-en-v2.0-hplt_opus
(Marian transformer-base, CC-BY-4.0) converted to CTranslate2 int8 for on-device use in the Mova
offline translation app. Weights are unchanged from upstream; only the format was converted
(the id-ordered .vocab was regenerated as the Marian YAML vocabulary the converter expects —
verified token-for-token against the SentencePiece model).
Quick facts
- ~81 MB on disk, ~110 MB resident RAM (CTranslate2 int8, CPU)
- Use with
ctranslate2.Translator+ the upstream SentencePiece model,beam_size=4
Evaluation note
We validated this conversion on a small corpus only — 60 FLEURS parallel sentences: chrF2 56.26 (uk→en), versus NLLB-600M-int8 58.10 on the same sentences. This is a sanity check that the conversion preserved quality, not a comprehensive benchmark. For proper evaluation numbers see the upstream HPLT v2.0 release and paper.
Attribution
- Model: HPLT project, CC-BY-4.0 — https://hplt-project.org
- Paper: An Expanded Massive Multilingual Dataset for High-Performance Language Technologies (https://arxiv.org/abs/2503.10267)
- Conversion: Mova project, 2026. License unchanged: CC-BY-4.0.