ckuma/lugbara-multilingual-translator

Use the fusion model instead. This XLM-R encoder-decoder (version 2, documented below) did not learn to translate on SALT: BLEU 0.00 (version 1 reached 0.29). For English to Lugbara and English to Luganda, use ckuma/lugbara-ensemble-translator, the XLM-R + mT5 fusion model (8.14 and 6.72 BLEU). For Lugbara to English and Luganda to English, use ckuma/lugbara-english-translator, the mT5 model (12.71 and 16.65 BLEU).

XLM-R English→Lugbara/Luganda translator

Fine-tuned FacebookAI/xlm-roberta-base for English → Lugbara, English → Luganda. It follows the model choice and training method of C. D. Akuma, An Optimized Machine Translation Model for Low-Resource Lugbara Language, MSc Computer Science, Muni University (October 2025): XLM-R was the thesis's optimised English→Lugbara model.

How to use

Put the target-language tag in front of the sentence and use lowercase text, as in training.

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tok = AutoTokenizer.from_pretrained("ckuma/lugbara-multilingual-translator")
model = AutoModelForSeq2SeqLM.from_pretrained("ckuma/lugbara-multilingual-translator")

def translate(text, tag):
    x = tok(tag + " " + text.lower(), return_tensors="pt")
    y = model.generate(**x, num_beams=4, max_new_tokens=128)
    return tok.decode(y[0], skip_special_tokens=True)

print(translate("women and girls' voices are often excluded from decision making.", "<2lgg>"))
print(translate("the cars have to be serviced before they go out to the field.", "<2lgg>"))

Tags: <2lgg> for English → Lugbara, <2lug> for English → Luganda.

Results (test split, 4656 sentence pairs)

Direction	Sentences	BLEU ↑	TER ↓	Word F1 ↑	chrF ↑	Apostrophe integrity ↑	Thesis best BLEU*
English → Lugbara	2328	0.0	99.94	0.0016	0.85	0.0035	22.18 (XLM-R (tuned))
English → Luganda	2328	0.0	99.88	0.0047	0.96	–	–

BLEU and TER are sacreBLEU corpus scores; word F1 is unigram precision/recall F1; apostrophe integrity is the share of apostrophe words in the reference ('ba, a'dule …) reproduced intact. *The thesis scores were measured on the thesis's own corpus, not on SALT, so they are a reference point rather than a like-for-like comparison.

Sample test translations (English → Lugbara)

Source	Reference	Model
women and girls' voices are often excluded from decision making.	ba erini tutuni o'duko oku eyini azini ezoanzi eyini ri e'yo peta pezu 'diyi ma alea ku
the cars have to be serviced before they go out to the field.	ba ma 'ye mutukari ma rua e'deta eri 'bazu pa dria 'diyi ra ka de nga fu ni mu amve azia ku
some of the government officials in uganda are just appointed into office.	ba lu azi'ba gameteni kari uganda niri ma alea 'diyi ma azi eyini azi fefe	diyi
the personal protective equipment is heavy and hot.	bongo azi'ba yi ageyi tepiri yini suu ri nzi nzi dika dri dri.	diyi
the task force managed to contain the spread of diseases in the district.	drice'ba azo korona niri ni 'diyi o'bi azo eyi ma ereta distrikitia ri atri ra
the area is out of bounds.	pari ndea 'ba leni 'ba ma ca ku
what could be the hindrances to the future development of youths?	e'yo ecopi 'ba ode eyi ma zota ewu drialeri ma alea ri atri pi ra 'diyi a'duni?	e
my grandfather died when he was eighty eight years old.	ma a'bi dra de eri ma eli kali aro drini aro.

Training

• Data: the thesis corpus (15,524 English–Lugbara pairs) is not public, so the same number of sentences was drawn at random (seed 13) from the cleaned Sunbird SALT text-all table and split 70/15/15 like the thesis: 10,867 train, 2,329 validation, 2,328 test. Luganda uses the same sentences. SALT has no domain labels, so the split is random rather than stratified by domain.

• Preprocessing (thesis §3.2.2): lowercased, punctuation normalised, all apostrophe variants folded to ', empty / duplicate / malformed pairs removed.

• Model: encoder-decoder warm-started from XLM-R base (Rothe et al., 2020), encoder and decoder weights shared.

• Hyperparameters (thesis §3.4.3): grid search over learning rate × batch size (3e-05/16 → 0.0 BLEU, 3e-05/32 → 0.07 BLEU, 5e-05/16 → 0.04 BLEU, 5e-05/32 → 0.1 BLEU, 0.0001/16 → 0.01 BLEU, 0.0001/32 → 0.06 BLEU), then fine-tuning with lr 5e-05, batch 32, optimiser adamw_torch, early stopping on validation loss (patience 5); 4,080 steps, 34.2 minutes on a Kaggle GPU.

• Version 2: grid settings compared after 2 passes instead of 1; the cross-attention layers (new, randomly initialised) learn at 10x the grid learning rate; label smoothing 0.1; early-stopping patience 5; repeated 3-word phrases are blocked when generating. Version 1 (1-pass grid, patience 3, no repeat blocking) stopped at validation loss 4.46 and scored BLEU 0.29 / 0.16 because its output looped.

See LUGBARA_ANALYSIS.md for the Lugbara language analysis of the training data.

Limits

• Trained on short SALT sentences (agriculture, health, community life); other domains are weaker.
• Lowercase only. Tone is not written in Lugbara, so tone-only differences are lost.
• Automatic scores use one reference per sentence; Lugbara speakers should review output before use.

Credits

• Thesis and method: Charity Driciru Akuma, Muni University (2025).
• Data: Sunbird AI, SALT (CC BY-SA 4.0).
• Base model: FacebookAI/xlm-roberta-base.
• Earlier work: mmuniversity/rutooro-multilingual-translator (Mountains of the Moon University).