A comprehensive guide to building a neural machine translation system for Dongxiang, a low-resource minority language from China. The tutorial covers the complete pipeline: preprocessing bilingual datasets, preparing tokenizers, registering new language IDs in NLLB-200, fine-tuning with Adafactor optimizer on 42,868 sentence pairs, evaluating with BLEU and ChrF++ metrics, and deploying via Hugging Face and GitHub Pages. The resulting models achieve BLEU-4 scores above 44 in both translation directions, demonstrating viable performance despite limited training data.
Table of contents
Background: Dongxiang as a Low-Resource LanguageOur Model: How to Use the Translation SystemModel Training: Step-by-Step Reproducible PipelineReflectionContact197 Impressions