Machine Translation: NLLB-200 Nepali Honorifics

Developed an honorific-aware English-to-Nepali Neural Machine Translation system using Meta’s NLLB-200 distilled 600M transformer model. Fine-tuned the pretrained model on a custom 25,343-pair English–Nepali parallel dataset, with emphasis on Nepali honorific levels including informal, neutral, and respectful forms.

Designed the training and evaluation pipeline using transformer-based sequence-to-sequence learning and attention mechanisms to improve contextual translation and honorific selection. Developed a Flask-based translation API to serve the fine-tuned model and tested its performance through automated and qualitative evaluation.

The model achieved 48.93 BLEU and 72.37 chrF++ on the evaluation dataset, demonstrating improved translation quality for the targeted honorific domain. The project was completed as a 7th semester academic research/project work, combining NLP, deep learning, model fine-tuning, dataset preparation, and API development.