galsenai/wolofToFrenchTranslator_nllb

🌍 Wolof ↔ French Translator (NLLB-based) 🇸🇳🇫🇷

This is a Wolof ↔ French translation model based on Meta AI’s NLLB (No Language Left Behind) architecture, fine-tuned specifically for high-quality bilingual translation in both directions.

🧠 Developed by GalsenAI, a Senegal-based open initiative promoting artificial intelligence for African languages.

---

🧠 About the Model

• Base architecture: facebook/nllb-200-distilled-600M
• Supported languages: Wolof (wo) ↔ French (fr)
• Purpose: To enable reliable translation for real-world applications like education, healthcare, and public services.
• BLEU score: 13 (on a custom Wolof-French evaluation set)

---

🚀 How to Use

📦 Install dependencies

pip install transformers torch

🔍 Example usage

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
import torch

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
checkpoint = "galsenai/wolofToFrenchTranslator_nllb"

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForSeq2SeqLM.from_pretrained(checkpoint).to(device)

def predict(text, lang):
    if lang.lower() == "wo":
        prefix = "translate Wolof to French: "
    elif lang.lower() == "fr":
        prefix = "translate French to Wolof: "
    else:
        raise ValueError("Invalid language code")
    inputs = tokenizer(prefix + text, return_tensors="pt").to(device)
    translated_tokens = model.generate(**inputs, max_length=30)
    return tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]

# Example
result = predict("Naka nga def?", lang="wo")
print(result) # "Comment ça va?"

---

📊 Performance

Translation Direction	BLEU Score
Wolof → French	13
French → Wolof	13

Note: BLEU score is an indicator of translation quality. Further training will improve results.

---

📚 Training Data

The model was fine-tuned using a mix of:

• Manually aligned Wolof–French parallel corpora
• Public resources (Common Voice, Wikipedia, administrative documents, etc.)
• Custom datasets collected via LinguaSprint Africa, a crowdsourcing platform for African languages.

---

🤝 Contributing

This model is maintained by GalsenAI.

If you’d like to:

• Help improve this model
• Contribute more Wolof/French data
• Build NLP tools for African languages

👉 Join us at github.com/GalsenAI or reach out to the team!

---

📜 License

MIT License — free to use for research, education, and social applications. 📣 Attribution requested: GalsenAI (2025)