from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

repo = "iranzi/lughalink-nllb-psa-en-kik"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSeq2SeqLM.from_pretrained(repo)
tok.src_lang = "eng_Latn"
text = "The public is advised to follow official health guidelines."
inputs = tok(text, return_tensors="pt")
bos = tok.convert_tokens_to_ids("kik_Latn")
out = model.generate(**inputs, forced_bos_token_id=bos, max_new_tokens=128)
print(tok.decode(out[0], skip_special_tokens=True))

Or via the LughaLink FastAPI Space / local API (POST /translate).

Training summary

Base: facebook/nllb-200-distilled-600M
Epochs: 1 (PSA silver bitext)
Platform: Navon / Kinesis A100
Report: https://github.com/unimart831-ai/lughalinkai/blob/main/DOCS/MODELLING_AND_TRAINING_REPORT.md

Limitations

Silver-trained; may copy teacher errors.
Limited domain outside PSA / public-notice style.
Cultural appropriateness needs human review.