Sunbird/translate-nllb-1.3b-salt

Model details

This machine translation model can convert single sentences from and to any combination of the following languages:

ISO 693-3	Language name
eng	English
ach	Acholi
lgg	Lugbara
lug	Luganda
nyn	Runyankole
teo	Ateso

It was trained on the SALT dataset and a variety of additional external data resources, including back-translated news articles, FLORES-200, MT560 and LAFAND-MT. The base model was facebok/nllb-200-1.3B, with tokens adapted to add support for languages not originally included.

Usage example

tokenizer = transformers.NllbTokenizer.from_pretrained(
    'Sunbird/translate-nllb-1.3b-salt')
model = transformers.M2M100ForConditionalGeneration.from_pretrained(
    'Sunbird/translate-nllb-1.3b-salt')

text = 'Where is the hospital?'
source_language = 'eng'
target_language = 'lug'

language_tokens = {
    'eng': 256047,
    'ach': 256111,
    'lgg': 256008,
    'lug': 256110,
    'nyn': 256002,
    'teo': 256006,
}

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
inputs = tokenizer(text, return_tensors="pt").to(device)
inputs['input_ids'][0][0] = language_tokens[source_language]
translated_tokens = model.to(device).generate(
    **inputs,
    forced_bos_token_id=language_tokens[target_language],
    max_length=100,
    num_beams=5,
)

result = tokenizer.batch_decode(
    translated_tokens, skip_special_tokens=True)[0]
# Eddwaliro liri ludda wa?

Evaluation metrics

Results on salt-dev:

Source language	Target language	BLEU
ach	eng	28.371
lgg	eng	30.45
lug	eng	41.978
nyn	eng	32.296
teo	eng	30.422
eng	ach	20.972
eng	lgg	22.362
eng	lug	30.359
eng	nyn	15.305
eng	teo	21.391