ckuma/lugbara-ensemble-translator

XLM-R + mT5 fusion translator (English ↔ Lugbara, English ↔ Luganda)

An ensemble of the two models of C. D. Akuma, An Optimized Machine Translation Model for Low-Resource Lugbara Language, MSc Computer Science, Muni University (October 2025). The thesis found XLM-R best for English→Lugbara and mT5 best for Lugbara→English; this model fuses them into one network that translates in all four directions.

source ──► XLM-R encoder ──► projection + LayerNorm ─┐
      └──► mT5 encoder ──────────────────────────────┴─► joint memory ──► mT5 decoder ──► translation

The mT5 decoder attends over the states of both encoders at every output word; a learned source-type vector tells it which encoder each state came from (dual-encoder fusion, cf. Zhu et al., 2020, Incorporating BERT into Neural Machine Translation).

How to use

The repository holds the two encoders' weights (mt5/, xlmr/), the fusion layers (fusion_head.safetensors) and the model code (fusion_mt.py, lugbara_mt.py).

import sys
from huggingface_hub import snapshot_download
d = snapshot_download("ckuma/lugbara-ensemble-translator")
sys.path.insert(0, d)
import fusion_mt

tm, tx, model = fusion_mt.build(from_dir=d)
print(fusion_mt.translate(tm, tx, model, ["<2lgg> farmers should plant early when the rains begin."]))
print(fusion_mt.translate(tm, tx, model, ["<2en> " + "<a lugbara sentence, lowercase>"]))

Tags: <2lgg> → Lugbara, <2lug> → Luganda, <2en> → English. Use lowercase input, as in training.

Results (test split, 9312 sentence pairs)

Direction	Sentences	BLEU ↑	TER ↓	Word F1 ↑	chrF ↑	Apostrophe integrity ↑	BLEU of the single models	Thesis best BLEU*
English → Lugbara	2328	8.14	93.79	0.3536	35.22	0.5545	0.0 (XLM-R alone)	22.18 (XLM-R (tuned))
English → Luganda	2328	6.72	88.93	0.277	36.52	–	0.0 (XLM-R alone)	–
Lugbara → English	2328	5.72	92.17	0.2566	26.53	–	12.71 (mT5 alone)	25.41 (mT5 (tuned))
Luganda → English	2328	6.57	89.84	0.2736	27.27	–	16.65 (mT5 alone)	–

The single models were trained and scored on exactly the same split. *Thesis scores were measured on the thesis's own corpus, not on SALT, so they are a reference point only.

Sample test translations

English → Lugbara

Source	Reference	Fusion 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	ba e'yo oku eyini 'diyi tuni e'yo 'bani 'ba eyini 'diyi ma dria
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	ba ma osu mutukari eyi nga po'di angu 'da ma alea 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	ba azi gamete ni 'diyi ma azi kari uganda niri ma alea 'diyi ma alea
the personal protective equipment is heavy and hot.	bongo azi'ba yi ageyi tepiri yini suu ri nzi nzi dika dri dri.	afa 'bani o'bi yini 'diyi ma rua 'diyi 'ba e'yo 'bazu 'diyi pie
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	ba e'yo 'ba e'yo 'bani 'ba e'yo 'da ma aza kozu distrikitia 'diyi ma alea ra

English → Luganda

Source	Reference	Fusion model
women and girls' voices are often excluded from decision making.	amaloboozi g'abakyala n'abaana ab'obuwala emirundi egisinga gaboolebwa mu kukola okusalawo.	abantu abawala n'abawala n'abawala n'abawala ebikwata ku bawala.
the cars have to be serviced before they go out to the field.	emmotoka zirina okukolebwa saaviisi nga tezinnatwalibwa kukozesebwa mu mirimu.	emmotoka erina okuteekerateekeddwa nga tebannattikirwa mu nkambi.
some of the government officials in uganda are just appointed into office.	abakungu ba gavumenti abamu mu uganda baweebwa buweebwa yafesi.	abakungu ba gavumenti aba gavumenti ba gavumenti abapya.
the personal protective equipment is heavy and hot.	ebintu ebikozesebwa okwetangira bizito, byokya era bitaataaganya.	ebintu by'obuyinza eby'obuyonjo eby'obuyonjo.
the task force managed to contain the spread of diseases in the district.	akakiiko akalabirizi kaasobola okufufugaza endwadde obutasaasaanira mu disitulikiti.	akakiiko akaddukanya ekirwadde kya disitulikiti mu disitulikiti.

Lugbara → English

Source	Reference	Fusion model
ba erini tutuni o'duko oku eyini azini ezoanzi eyini ri e'yo peta pezu 'diyi ma alea ku	women and girls' voices are often excluded from decision making.	pregnant women should be taught how to avoid hiv and stay healthy.
ba ma 'ye mutukari ma rua e'deta eri 'bazu pa dria 'diyi ra ka de nga fu ni mu amve azia ku	the cars have to be serviced before they go out to the field.	motorcyclists are not allowed to move out of their cars without contamination.
ba lu azi'ba gameteni kari uganda niri ma alea 'diyi ma azi eyini azi fefe	some of the government officials in uganda are just appointed into office.	some ugandans are found guilty of corruption.
bongo azi'ba yi ageyi tepiri yini suu ri nzi nzi dika dri dri.	the personal protective equipment is heavy and hot.	employers are not following the rules and regulations.
drice'ba azo korona niri ni 'diyi o'bi azo eyi ma ereta distrikitia ri atri ra	the task force managed to contain the spread of diseases in the district.	the number of coronavirus cases in the district is worrying.

Luganda → English

Source	Reference	Fusion model
amaloboozi g'abakyala n'abaana ab'obuwala emirundi egisinga gaboolebwa mu kukola okusalawo.	women and girls' voices are often excluded from decision making.	pregnant mothers and childbirth are essential for expectant mothers.
emmotoka zirina okukolebwa saaviisi nga tezinnatwalibwa kukozesebwa mu mirimu.	the cars have to be serviced before they go out to the field.	the vehicle that knocked him down should be checked for further investigations.
abakungu ba gavumenti abamu mu uganda baweebwa buweebwa yafesi.	some of the government officials in uganda are just appointed into office.	some government officials have been suspended in uganda.
ebintu ebikozesebwa okwetangira bizito, byokya era bitaataaganya.	the personal protective equipment is heavy and hot.	according to the findings, the truth eventually, the problem is a recurring theme.
akakiiko akalabirizi kaasobola okufufugaza endwadde obutasaasaanira mu disitulikiti.	the task force managed to contain the spread of diseases in the district.	the task force has to stop the spread of diseases in the district.

Training

• Data: 10,867 / 2,329 / 2,328 sentences (train / validation / test) drawn at random (seed 13) from Sunbird SALT, the thesis's 15,524-sentence, 70/15/15 design; every sentence is used in all four directions.
• Learning rates: mT5 and the fusion layers 0.001 (mT5's grid-search winner); XLM-R 3e-05 (XLM-R's grid-search winner); Adafactor; XLM-R's word embeddings frozen.
• Training: batch 16, early stopping on validation loss (patience 3), 19,019 steps, 288.9 minutes on a Kaggle T4; 860M parameters (668M trained).

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

Limits

• Short SALT sentences (agriculture, health, community life); other domains are weaker.
• Lowercase only; Lugbara tone is not written, 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 models: google/mt5-base, FacebookAI/xlm-roberta-base · Earlier work: mmuniversity/rutooro-multilingual-translator.