CoolFace
Modelpublic

iranzi/lughalink-nllb-psa-en-sw

sourceHugging Facemitupdated 2mo agoView on Hugging Face
0likes21downloads
Model Card

LughaLink NLLB PSA — en-sw

Fine-tuned NLLB-200 distilled 600M for English → Kiswahili Public Service Announcements (Kenya).

Intended use

  • —Translate short English PSA / public advisory text into Kiswahili.
  • —Research and course demo (DSA 4020); not a certified government translation service.

Training data (honesty)

  • —Source: framework-filtered Kenyan PSA English (real + labeled synthetic).
  • —Targets: silver NLLB zero-shot seeds (verified=false), auto-QC filtered.
  • —Not human gold. Do not treat outputs as officially verified.

Languages

RoleLanguageNLLB code
SourceEnglisheng_Latn
TargetKiswahiliswh_Latn

How to use

python
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

repo = "iranzi/lughalink-nllb-psa-en-sw"
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("swh_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/MODELLINGANDTRAINING_REPORT.md

Limitations

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