Nvidia and IIT Madras research lab AI4Bharat are preparing an open training resource containing about 2.8 trillion tokens across 22 Indian languages. The Nemotron-pretraining-Indic-v1 release will include native-script material for all 22 languages and romanized text for 12, producing 34 language-and-script configurations.
The collaborators also plan to release the GPU-accelerated pipeline used to build the corpus. It covers document translation, quality assessment and targeted repair, including the prompts and language-model workflows behind those stages. Every generated record is linked to a curated source document and checked against its input, according to the project team.
The material draws from English sources that include mathematics, code and question-answer content. Its purpose is to give model builders more data for systems that explain technical ideas, solve problems and answer questions in Indian languages, where large pretraining collections remain less available than in English.
The corpus and pipeline are described as releasing soon, so developers cannot yet treat the announcement as a completed dataset delivery. Synthetic translation can also carry errors or cultural distortions despite automated checks. The open recipe will matter because it lets researchers inspect the process, regenerate material and test quality for their own languages and applications.