NVIDIA and IIT Madras release 2.8T-token Indic pretraining corpus and open pipeline

NVIDIA, with IIT Madras and AI4Bharat, introduced Nemotron-pretraining-Indic-v1, a synthetic pretraining corpus of about 2.8 trillion tokens across 22 Indian languages, plus the open GPU-accelerated generation pipeline. The corpus holds 2,601.0 billion native-script tokens and 224.6 billion romanized tokens across 34 language-script configurations.

• Corpus: 2,601.0B native-script tokens and 224.6B romanized tokens
• 22 languages are India's constitutional languages, first languages for ~1.2B people
• Translation used Gemma-4-31B-it with a 75-out-of-100 quality threshold
• NeMo Data Designer pipeline is open, with every output traced to its source