0x0checo/Luganda-to-English-machine-translation

# Fine-tuning Multilingual NMT for a Low-Resource, Morphologically Rich Language (Luganda → English) This project studies **Luganda–English machine translation** under low-resource conditions. We evaluate several baselines (including **Google Translate**) and fine-tune two multilingual pretrained NMT models—**NLLB-600M** and **mBART-50**—on the **SALT Luganda–English** parallel corpus. ## Dataset We use the **SALT corpus** (≈ **25,000** Luganda–English sentence pairs), split into: * Train: **23,497** * Validation: **496** * Test: **500** ## Models & Setup **Baselines** * Copy baseline (returns the source as output) * mBART-50 baseline (uses **Swahili `sw_KE`** as `src_lang` since Luganda is not in pretraining) * NLLB-600M baseline (`lug_Latn` → `eng_Latn`) * Google Translate **Fine-tuning** * Fine-tuned mBART-50 (training stopped after **2 epochs** due to disk quota limits) * Fine-tuned NLLB-600M * Fine-tuned NLLB-600M + **Back-Translation** (train en→lug, synthesize Luganda, then retrain lug→en) Evaluation uses **BLEU** on the same test set with consistent decoding settings. ## Results (BLEU) | Model | BLEU | | | --------------------------------------- | --------: | - | | Copy Baseline | 2.84 | | | Google Translate | 9.24 | | | mBART-50 Baseline | 7.60 | | | NLLB-600M Baseline | 15.62 | | | Fine-tuned mBART-50 | 25.93 | | | Fine-tuned NLLB-600M | **35.99** | | | Fine-tuned NLLB-600M + Back-Translation | 35.99 | | **Key takeaway:** fine-tuning yields large gains, and **NLLB-600M** performs best; back-translation did not further improve BLEU in this setup. ## Acknowledgements * SALT dataset: Nabende et al. (2023) * Models: mBART (Liu et al., 2020), NLLB (Team et al., 2022)