Sinhala–Tamil NMT

// Overview

End-to-end neural machine translation research project targeting the low-resource Sinhala–Tamil language pair. Built on mBART-50 with custom fine-tuning on approximately 25,000 parallel sentence pairs. Evaluated against the FLORES-200 benchmark using a reproducible three-seed BLEU/chrF protocol. Includes a FastAPI inference server and a Next.js prototype UI for live translation.

// Gallery

Bridging Sinhala and Tamil through fine-tuned mBART-50 neural machine translation

FLORES-200 benchmark: fine-tuned model vs baseline across three training seeds

Next.js prototype UI: live translation with 87.3% confidence and 245 ms latency monitoring

// Key Achievements

• Fine-tuned mBART-50 on ~25K Sinhala–Tamil parallel sentence pairs
• FLORES-200 external test set evaluation for reproducible benchmarking
• Three-seed BLEU/chrF evaluation protocol to control for training variance
• FastAPI inference server with batch translation and confidence scoring
• Next.js prototype UI for live translation with latency monitoring

// Stack