NVIDIA, IIT Madras and its AI4Bharat lab have introduced Nemotron-Pretraining-Indic-v1, a synthetic pretraining corpus of roughly 2.8 trillion tokens covering 22 Indian languages. Why it matters: a lack of training data is the main reason Indian-language AI lags English — this is one of the largest open attempts yet to close that gap, and the recipe for making more of it is open too.
October 15, 2026 — The corpus and its generation pipeline were introduced in a Hugging Face post by AI4Bharat and NVIDIA researchers.
What is in the Indic corpus?
The 22 languages are the ones listed in India’s Constitution, described in the announcement as the first languages of about 97% of Indians — nearly 1.2 billion people at the 2011 Census. Token counts, measured with the Gemma-4 tokenizer, break down as about 2,601 billion native-script tokens and about 225 billion romanized tokens, across 34 language-and-script configurations. All 22 languages get native-script data; twelve also get romanized versions, reflecting how many people actually write those languages in Latin script.
The corpus is described as releasing soon, alongside the open, GPU-accelerated pipeline used to build it.
How was the data made?
The teams started from English documents in 15 subsets of Nemotron-CC-v2 — web text, question-and-answer material, mathematics, code and other specialised content — and translated them at document level. For twelve languages, translation and quality scoring used Gemma-4-31B-it; the other ten used an Indic Translate workflow built for English and 22 Indian languages. Outputs had to score at least 75 out of 100 on a quality rubric covering meaning, fluency, terminology, style and formatting. Lower-scoring translations went through a guided repair step, which the teams say recovered about 438 billion tokens that are included in the total. Every record keeps a link back to its source document.
Has the corpus already improved a model?
Yes, according to the teams. A 10-billion-token slice of Indic data formed part of a 100-billion-token continued-pretraining stage for Nemotron 3.5 Super VL. The announcement says scores rose on all five reported Indic knowledge and reasoning benchmarks, by roughly 5 to 10 percentage points — while noting that the stage mixed synthetic and natural Indic text, so the gain cannot be credited to this corpus alone.
Why it matters
Most open training data is overwhelmingly English, so models serving India’s languages start at a disadvantage. A corpus at this scale, with traceable sources and an open generation recipe that other teams can adapt to their own documents, gives model builders a realistic way to do better. The limits are stated plainly in the announcement itself: translated text does not replace material originally written in Indian languages, and quality checking was mostly automated with sampled human review. Used alongside natural Indic text, though, this is a serious new building block — everything AI, tested, starts with what the model was trained on.
Frequently asked questions
Can I download Nemotron-Pretraining-Indic-v1 now?
The announcement says the corpus is releasing soon. The open generation pipeline — built on NVIDIA’s NeMo Data Designer — is shared so teams can also create their own data.
Which languages does it cover?
All 22 languages listed in India’s Constitution, in native script, with romanized versions for twelve of them.
Is the data human-written?
No. It is synthetic data: English source documents translated and quality-checked by models, with each record traceable to its source. The creators recommend combining it with naturally written Indic text.
Sources: Hugging Face (NVIDIA and AI4Bharat), Unite.AI