nllb-kirundi-multi (Umuseke w'Ijwi) 🇧🇮

Ikirundi cacu, Ijwi ryacu!
(Our language, our voice!)

Fine-tuned NLLB-200-600M model for multidirectional translation.
Part of the Umuseke w'Ijwi initiative for African Linguistic Sovereignty.

---

🌍 Vision

Our mission is to bridge the "Digital Linguistic Divide" for the 30 million Kirundi speakers in East Africa. By creating open-source AI infrastructure, we ensure that Kirundi is a first-class citizen in the global digital revolution.

This model serves as a foundational "engine" for:

• 🎙️ Speech-to-Text (ASR)
• 🗣️ Text-to-Speech (TTS)
• 🌐 Machine Translation (MT)

🚀 Key Features

This is a multidirectional translation model (Seq2Seq) capable of translating:

• Kirundi (run_Latn) ↔ French (fra_Latn)
• Kirundi (run_Latn) ↔ English (eng_Latn)

📊 Training & Performance

• Base Model: Meta NLLB-200-distilled-600M.
• Dataset: Kirundi Open Speech Dataset (approx. 11k pairs in this version).
• Current Score: 9.94 BLEU (Iteration V1). Performance is expected to increase as the community adds more validated "Gold" phrases.

🛠️ Usage (Python)

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("Ijwi-ry-Ikirundi-AI/nllb-kirundi-multi")
model = AutoModelForSeq2SeqLM.from_pretrained("Ijwi-ry-Ikirundi-AI/nllb-kirundi-multi")

# Example: Kirundi -> French
tokenizer.src_lang = "run_Latn"
text = "Amahoro y'Imana abane nawe."

inputs = tokenizer(text, return_tensors="pt")
output = model.generate(
    **inputs, 
    forced_bos_token_id=tokenizer.convert_tokens_to_ids("fra_Latn")
)

print(tokenizer.decode(output[0], skip_special_tokens=True))

🫱🏽‍🫲🏻 Community & Open Source

This project is dedicated to the Open Source community. We believe that linguistic data and AI models should be public goods to foster innovation and inclusion across Africa.

---

👥 Contributors

• Samandari: github.com/Sama-ndari
• Rugwe: github.com/Irumva17

Built by Ijwi ry'Ikirundi AI.