IndicTrans2

IndicTrans2 is the first open-source transformer-based multilingual neural machine translation model covering all 22 scheduled Indian languages, from AI4Bharat at IIT Madras. The paper (arXiv 2305.16307) dates to May 2023; exact release day not confirmed. It uses script unification to enable transfer learning through lexical sharing between languages, and supports five scripts: Perso-Arabic, Ol Chiki, Meitei, Latin, and Devanagari.

AI4Bharat open-sourced the full stack: the training dataset (BPCC), back-translation data (BPCC-BT), final models, the IN22 evaluation benchmark, and training and inference code. The models power national translation infrastructure, judicial multilingual tools, and civic technology in India, and a live demo runs at models.ai4bharat.org.