SsemuliJoseph/nllb-luganda-baseline

NLLB Luganda Baseline (English → Luganda)

Fine-tuned baseline translation model for English → Luganda, built on facebook/nllb-200-distilled-600M. Trained as part of an industrial-training NLP pipeline at Sunbird AI, targeting low-resource Ugandan language translation.

Model Description

• Base model: facebook/nllb-200-distilled-600M
• Architecture: M2M-100 family (NLLB), 0.6B parameters
• Task: Text2Text Generation — English to Luganda translation
• Language(s): English (en), Luganda (lg)
• Developed by: SsemuliJoseph
• Finetuned from model: facebook/nllb-200-distilled-600M

Training Data

Fine-tuned on a small, hand-curated golden evaluation set produced via an Argilla annotation pipeline (Accept / Needs Editing / Reject review discipline). This is a deliberately small baseline dataset — see "Limitations" below.

Evaluation

Model	BLEU
Baseline (this model)	17.84

Evaluated qualitatively and quantitatively against a held-out golden set (see companion LoRA-adapted model for comparison).

Intended Use

Research and prototyping for English–Luganda machine translation in low-resource settings. Demonstrated live in the luganda-translator-demo Space.

Limitations

• Trained on a small dataset; BLEU of 17.84 reflects early-stage baseline performance, not production quality.
• Not evaluated for bias, toxicity, or robustness across Luganda dialects.
• Should not be used for legal, medical, or safety-critical translation.

How to Use

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("SsemuliJoseph/nllb-luganda-baseline")
model = AutoModelForSeq2SeqLM.from_pretrained("SsemuliJoseph/nllb-luganda-baseline")

inputs = tokenizer("Hello, how are you?", return_tensors="pt")
outputs = model.generate(**inputs, forced_bos_token_id=tokenizer.convert_tokens_to_ids("lug_Latn"))
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Demo

Try it live →