NLLB-Tarjumi: Fine-tuned Translation for African Languages

Fine-tuned NLLB-200-distilled-600M on parallel corpus data for Kenyan languages.

Training Results

Language	Base BLEU	Fine-tuned BLEU	Improvement
Kikuyu (ki)	17.00	22.24	+31%
Luo (luo)	11.29	18.64	+65%
Swahili (sw)	16.45	19.39	+18%

Training Details

• Base model: facebook/nllb-200-distilled-600M
• Method: LoRA (rank-16, alpha-32)
• Corpus: 124K parallel pairs (Bible + JW300)
• Hardware: NVIDIA H100 80GB
• Training time: 13 minutes
• Cost: ~$0.75

Supported Languages

16 African languages: Kikuyu, Swahili, Luo, Luhya, Kalenjin, Kamba, Meru, Maasai, Somali, Amharic, Yoruba, Zulu, Xhosa, Shona, Kinyarwanda, Luganda.

Usage

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model = AutoModelForSeq2SeqLM.from_pretrained("cif-ai/nllb-tarjumi")
tokenizer = AutoTokenizer.from_pretrained("cif-ai/nllb-tarjumi")

tokenizer.src_lang = "eng_Latn"
inputs = tokenizer("Hello, how are you?", return_tensors="pt")
tgt_id = tokenizer.convert_tokens_to_ids("kik_Latn")

translated = model.generate(**inputs, forced_bos_token_id=tgt_id, max_new_tokens=128)
print(tokenizer.decode(translated[0], skip_special_tokens=True))

Product

This model powers Tarjumi — a free translation service for African mother-tongue languages.

Built by CIF AI.