NVIDIA opens 2.83-trillion-token Indic AI corpus
NVIDIA and AI4Bharat are opening a 2.83-trillion-token Indic training corpus and the recipe used to generate it.
NVIDIA and AI4Bharat have unveiled an open Indic-language pretraining corpus containing 2.83 trillion target tokens across 22 languages, alongside the data-generation workflow used to build it. The project is aimed at giving model builders more training material for Indian languages and a reproducible way to make more. 1
What NVIDIA is releasing
The Nemotron-pretraining-Indic-v1 corpus combines native-script and romanised data. The project describes 34 language-script configurations and 2.44 billion document records, with the data derived from curated English documents. The 22 languages are those listed in India’s Constitution, and the partners say they cover the first languages of roughly 97 per cent of the country’s population. 1
The release is more than a large token count. NVIDIA, AI4Bharat and IIT Madras are also opening the generation recipe, including prompts, model workflows and quality checks. Developers can inspect the process, adapt it to their own source material and trace generated records back to source documents. 1
Why the data matters
AI systems have historically had much more English training data than material for Indian languages. The new corpus attempts to close part of that gap by translating curated English content covering mathematics, code and question-and-answer material, then checking outputs against their source documents. The project says accepted translations are filtered through automated quality assessment, with lower-scoring outputs repaired and assessed again. 1
The dataset includes 2,273.8 billion native-script tokens across 12 languages, 224.6 billion romanised tokens for those same languages and another 327.2 billion native-script tokens across 10 more languages. The accompanying composition chart makes clear that these are corpus-inventory figures, not a claim that every token represents a unique English source document. 1
Model connection
The announcement says continued pretraining for Nemotron 3.5 Super VL added 10 billion Indic tokens from a 12-language subset of the corpus and natural Indic text. NVIDIA reports gains across five Indic knowledge and reasoning benchmarks after that stage, but the comparison covers a combined synthetic-and-natural recipe rather than isolating this corpus alone. 1
Our opinion
The useful part of this announcement is the open recipe as much as the headline token count. More data can improve language coverage, but publishing the prompts, checks and provenance gives researchers a better chance of finding weaknesses rather than treating a giant dataset as self-validating. The project is promising infrastructure, not proof that Indian-language AI quality is solved: native-speaker evaluation, licensing checks and tests for residual translation errors will matter as much as scale.
For developers, the immediate story is access to a larger Indic training resource and a documented method for generating more. For everyone else, the important caveat is that the reported benchmark gains belong to a wider continued-pretraining stage and should not be read as a clean measurement of this corpus by itself.