opus-mt-en-mk-ct2

opus-mt-en-mk-ct2

English → Macedonian, converted to CTranslate2 and quantised to int8 for on-device translation in Kiwaro.

Converted from Helsinki-NLP/opus-mt-en-mk.

Measured

12 parallel FLEURS sentences, EN→MK, beam size 4, int8 on CPU:

Twelve sentences is a small sample: it establishes that the model works, not precisely how well. For comparison, the Macedonian→English direction scores chrF 69.9 using opus-mt-tc-big-zls-en.

Files

The .vocab files are generated from the .spm files and ship inside the app rather than being downloaded — the app's tokeniser is Swift and never sees a .spm. Both are published so the pack can be rebuilt without repeating the conversion.

Licence and attribution

Apache 2.0, which is the licence of the source model — not CC BY 4.0.

This matters because Kiwaro's other OPUS-MT pack is different: Helsinki-NLP/opus-mt-tc-big-zls-en, used for Macedonian→English, is CC BY 4.0. The two packs are not under the same terms and the app's attribution notice has to say so. Checked against the Hugging Face API on 4 September 2026 rather than assumed — the first draft of this card said CC BY 4.0 by copying the other pack, which would have been a false licence statement on a public repository.

What Apache 2.0 asks of us

• Keep the licence and copyright notices (this file, plus LICENSE alongside it).
• State what changed: converted to CTranslate2 and quantised to int8. No retraining, no fine-tuning, no change to the weights beyond quantisation.
• No attribution requirement in the CC BY sense, but the original work is credited below because it should be.

Original model by the Language Technology Research Group at the University of Helsinki. Please cite the OPUS-MT work: Tiedemann & Thottingal, OPUS-MT — Building open translation services for the World, EAMT 2020.