Fine-tuned from facebook/nllb-200-distilled-600M on East African AML transaction narratives. Specialized for translating Luganda (lug_Latn) and Swahili (swh_Latn) mobile money transaction descriptions to English for downstream AML classification.

Why a specialized translator?

General NLLB models miss domain-specific AML vocabulary:

Mobile money agent terminology (float, airtime, USSD codes)
Ugandan colloquialisms used in social engineering scams
Financial crime typology phrases specific to EAC corridor

Usage

python

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_id = "darthvader256/simitech-aml-afrinllb-translator"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

tokenizer.src_lang = "lug_Latn"
inputs = tokenizer("nkusaba ssente z'omusawo omukisa", return_tensors="pt")
output = model.generate(
    **inputs,
    forced_bos_token_id=tokenizer.lang_code_to_id["eng_Latn"],
    max_new_tokens=128,
)
print(tokenizer.decode(output[0], skip_special_tokens=True))
# → "I am asking for doctor money, please"

Source

decision-plane/app/training/nlp_finetune.py — AfriNLLBTranslator class