Manipuri ↔ English Translator

The problem

Manipuri has its own script and very little parallel data, so off-the-shelf translation models do not cover it.

What made it hard

No pretrained checkpoint exists for the pair — the tokenizer and the model both had to be trained from nothing.

Architecture

1. 01Corpus assembly from the sparse available parallel text
2. 02Subword tokenizer trained for the non-Latin script
3. 03Encoder-decoder transformer trained from scratch in PyTorch
4. 04BLEU evaluation against the baselines that exist

Outcomes

• Working translation for a pair with no pretrained model available
• Documented end-to-end in a reproducible notebook