Mateza Marian en↔rw: English ↔ Kinyarwanda translation

One model, both directions. A 61M-parameter Marian Transformer. Two tags at the start of the input, <src_en><tgt_rw> or <src_rw><tgt_en>, pick the direction.

Built by Promesse Rukundo and Mateza Kayondo (Mateza).

Quick start

pip install torch "transformers>=5.0" huggingface_hub

import os, sys
from huggingface_hub import hf_hub_download
sys.path.insert(0, os.path.dirname(hf_hub_download("mateza/mateza-marian-en-rw", "translator.py")))
from translator import Translator

t = Translator("mateza/mateza-marian-en-rw")               # GPU (fp16) if available
t.translate("Good morning, how are you?", "en->rw")        # 'Mwaramutse, amakuru?'
t.translate(["Murakoze cyane.", "Abana bagomba kujya ku ishuri buri munsi."], "rw->en")
# ['Thank you so much.', 'Children should go to school every day.']
t.translate("Wash your hands with soap.", "en->rw", beams=4)   # beam 4: about 0.5 chrF++ better, slower

Command line. Put the direction in quotes, because the shell treats a bare > as a redirect:

python translator.py --repo . --direction "en->rw" "Thank you very much."
python translator.py --repo . --direction "rw->en" --beams 4 "Murakoze cyane."

Plain transformers, without translator.py

The tokenizer does not add the language tags itself, so add them yourself:

import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

repo = "mateza/mateza-marian-en-rw"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSeq2SeqLM.from_pretrained(repo).eval()

def encode(text, src, tgt):
    tags = tok.convert_tokens_to_ids([f"<src_{src}>", f"<tgt_{tgt}>"])
    return tags + tok(text, add_special_tokens=False).input_ids + [tok.eos_token_id]

ids = torch.tensor([encode("Thank you very much.", "en", "rw")])
out = model.generate(input_ids=ids, num_beams=1, max_new_tokens=128, no_repeat_ngram_size=3, repetition_penalty=1.1)
print(tok.decode(out[0], skip_special_tokens=True))      # 'Urakoze cyane.'

Only en and rw were trained. The vocabulary also has tags for fr, ar, sw, lg and zh_Hans, which are reserved for later versions. Don't use them yet.

Evaluation

FLORES+ devtest (1,012 sentences per direction)

No empty outputs and no degenerate (repetitive) outputs. At least 99.8% of outputs are in the requested language.

How we got here (FLORES+ devtest, mean chrF++, beam 4)

Domain test sets (500 held-out sentences each, greedy chrF++)

These are the test splits of the datasets used for the fine-tune. No test sentence is in the training data, but the sentences come from the same sources, so these numbers are in-domain. FLORES+ is the fairer general measure.

Sample translations (greedy)

The full reports are in eval/:

• REPORT.md and summary.json: FLORES+ results;
• DOMAINS.md and domains.json: before and after, per domain;
• probes.tsv: the sample phrases;
• train-report.json and data-report.json: the training run and the data counts.

Limitations

• Clock times are unreliable. Rwandan hours count from 6 a.m. ("saa munani" = 2 p.m.). Check any time, date or number.
• Some greetings are still weak in rw→en. For example, "Muramuke" is translated as "Next".
• No native-speaker review yet. The scores are automatic metrics. Don't use the output unreviewed for medical, legal or other high-stakes text.
• Input length: up to 256 tokens per call. Translate long documents sentence by sentence.

Training

New data in stage 2

Every row went through the same checks:

• broken characters fixed or removed;
• untranslated rows dropped;
• length and length-ratio checks;
• markup and scripture filters;
• LaBSE alignment, at 0.70 for web-mined sources and 0.50 for human-translated ones (not applied to the curated packs);
• duplicates removed, including against the stage-1 data;
• anything that appears in FLORES+ or in a held-out test or dev split removed.

Left out on purpose:

• machine-translated sets;
• software interface strings (GNOME, KDE, Ubuntu);
• more web-mined NLLB;
• a scripture set with its stopwords stripped.

Data and attribution

The model weights are released under Apache-2.0. Training data, with the licences as stated by each source:

• Digital Umuganda:
  • Rwandan Gazette: CC-BY-4.0.
  • Health: no licence stated.
  • Kinyarwanda–English MT dataset: CC-BY-4.0.
• mbazaNLP: Education, CC-BY-2.0.
• Google: SMOL, CC-BY-4.0.
• TICO-19: gmnlp/tico19, CC0.
• OPUS:
  • Wikimedia: CC-BY-SA.
  • translatewiki: CC-BY.
  • Tatoeba: CC-BY 2.0 FR.
  • J. Tiedemann, Parallel Data, Tools and Interfaces in OPUS (LREC 2012).
• Kira-Floris: cross-lingual-translation-en-kin-parquet, expert_en_rw config; no licence stated.
• Hamed-PRO: interuro-kinyarwanda-sentences-english, MIT.
• NLLB: mined bitext (allenai/nllb), ODC-BY 1.0, via the v3 data mix.
• Mateza: Ice Block curated packs.
• Teachers: mateza/mateza-opus-ft-v3, which was fine-tuned from Helsinki-NLP OPUS-MT (Tiedemann and Thottingal, EAMT 2020).

Evaluation uses FLORES+ (CC-BY-SA-4.0), which isn't redistributed here.

Files

Citation

@misc{mateza-marian-en-rw,
  title  = {Mateza Marian en-rw: one-model English-Kinyarwanda translation},
  author = {Rukundo, Promesse and Kayondo, Mateza},
  year   = {2026},
  note   = {Distilled from Mateza OPUS-FT v3, then fine-tuned on Gazette, Health, Education and everyday data. Organization: Mateza.},
  url    = {https://huggingface.co/mateza/mateza-marian-en-rw}
}