AfriSpeech/afrigloss-opus-mt-fr-ha 🌍

AfriSpeech/afrigloss-opus-mt-fr-ha is a production glossary-constrained machine translation model for French $\rightarrow$ Hausa, fine-tuned under the Afri-Gloss-MT framework.

Part of the Afri-Gloss-MT Models Collection.

---

🌟 Key Capabilities

• Zero-Shot Terminology Control: The model accepts user-injected or dictionary-matched target terms in angle brackets (<target_term>) inside the source sentence, and reliably preserves them in the generated translation.
• Trained to Convergence on Full Bitext: Fine-tuned on full multi-version parallel corpora from AfriSpeech/africa-corpus and AfriSpeech/african-corpus-jw with 100% verified in-place noun-chunk replacements.
• High-Performance Inference: Built on MarianMT for fast, lightweight CPU and GPU deployment.

---

🚀 Quickstart & Inference

Python (Hugging Face transformers)

from transformers import MarianMTModel, MarianTokenizer

model_name = "AfriSpeech/afrigloss-opus-mt-fr-ha"
tokenizer = MarianTokenizer.from_pretrained(model_name)
model = MarianMTModel.from_pretrained(model_name)

# 1. Plain source sentence with in-place target glossary insertion
prompt = "The <malam> brought <salama> and <farin ciki> to the children."

inputs = tokenizer(prompt, return_tensors="pt", padding=True)
translated_tokens = model.generate(**inputs, max_length=128, num_beams=4)
translation = tokenizer.decode(translated_tokens[0], skip_special_tokens=True)

print("Output Translation:", translation)
# Output preserves the injected terms fluently!

---

📊 Dataset Reference

This model was trained on the hau subset of the official dataset:

• Dataset Repository: AfriSpeech/gloss-mt-train-fr
• Subset Name: "ha"

---

📚 Citation

@article{owusu2026afrigloss,
  title={Afri-Gloss-MT: In-Place Glossary Constraints for African Machine Translation},
  author={Owusu, Mich-Seth},
  journal={arXiv preprint},
  year={2026}
}