The short answer: NVIDIA and AI4Bharat at IIT Madras announced Nemotron-pretraining-Indic-v1 on 16 Oct 2026, a training dataset of about 2.8 trillion tokens covering 22 Indian languages. All of it is machine-translated from English. The dataset is “releasing soon”, and the pipeline used to create it is already open source.
Key facts
- What: Nemotron-pretraining-Indic-v1, a pretraining dataset for Indian languages.
- Who: NVIDIA and AI4Bharat at IIT Madras, with a translation model from Bodhan AI.
- Size: about 2.8 trillion tokens across 22 languages and 34 language-script combinations.
- How: machine-translated from English, with an AI judge scoring every translation.
- Status: dataset “releasing soon”; the generation pipeline is open source on GitHub.
Ask most AI chatbots something in English and you get a confident essay. Ask the same thing in Odia or Konkani and you often get something that reads like it was translated by a sleepy intern.
The reason is simple: data. Models learn languages from text, and the internet is mostly English.
Mitesh Khapra of AI4Bharat, the AI lab at IIT Madras, put it bluntly: “we do not have enough data for Indian languages.” Last week, NVIDIA and AI4Bharat published a big attempt to change that. (NVIDIA and AI4Bharat on Hugging Face)
What exactly did they release?
It is called Nemotron-pretraining-Indic-v1. It contains about 2.8 trillion tokens across 22 Indian languages, in 34 language-and-script combinations. That breaks down to roughly 2,601 billion tokens in native scripts and 225 billion in romanised text, the “Hinglish-style” typing most of us actually use on WhatsApp. (Hugging Face)
The 22 languages:
- Native script plus romanised (12): Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Tamil, Telugu, Urdu.
- Native script only (10): Assamese, Bodo, Dogri, Kashmiri, Konkani, Maithili, Manipuri, Sanskrit, Santali, Sindhi.
If your language is on the second list, this is a bigger deal than it sounds. These languages have very little text online for models to learn from.
How do you make 2.8 trillion tokens of Indian-language text?
You translate. All of the content is machine-translated from English, starting from 15 subsets of NVIDIA’s existing English web, Q&A, maths and code data. (Hugging Face)
For the 12 bigger languages, a model called Gemma-4-31B-it translates full documents and then grades its own homework, scoring each translation against the source. Anything below 75 out of 100 gets repaired and checked again. For the other 10 languages, Bodhan AI’s Indic Translate does the translating, with Gemma as the examiner. (Hugging Face)
Does it actually make AI better at Indian languages?
Early signs say yes. NVIDIA gave its Nemotron 3.5 Super VL model a small top-up of 10 billion Indic tokens. Scores on five Indian-language benchmarks, including MILU and Indic GSM8K, improved by between 4.98 and 9.82 percentage points. (Hugging Face)
It was not a clean sweep. One instruction-following benchmark, IFEval 1.5, dropped by about 2 points. Every gain has a receipt. (Hugging Face)
The catch (and it is a fair one)
Translated English is not the same as Indian writing. A translated article about American baseball, now in Tamil, teaches the grammar of Tamil. It does not teach how Chennai talks about cricket, festivals or rent.
The authors seem aware of this. They describe natural Indian-language text as a complement to the translated data, not something it replaces. (Hugging Face)
And a practical note: the dataset itself is listed as “releasing soon”, with no licence named yet. The recipe for generating it is already open on GitHub. (Hugging Face)
Our read (opinion)
The headline number is 2.8 trillion. The more important word is “recipe”. An open pipeline means Indian startups, universities and even state governments can generate data for their own domains, from agriculture advisories to court records, instead of waiting for Silicon Valley to care.
This is also the boring infrastructure that makes models like Monday’s Gemini 4 Argon useful to the next billion users. Frontier models get the headlines. Datasets decide who they work for.
One prediction: the next big Indian AI company will not win on model size. It will win on whose language it understands best.
Do this
Test your favourite AI tool in your mother tongue this week. Ask it one practical question (a recipe, a bank process, a school form) in native script, then in romanised text. Note where it stumbles. If you build products, that gap is your market. If you just use them, you now know when to switch back to English.
FAQ
What is Nemotron-pretraining-Indic-v1?
It is a pretraining dataset of about 2.8 trillion tokens covering 22 Indian languages, built by NVIDIA and AI4Bharat at IIT Madras and announced on 16 Oct 2026.
Which Indian languages does the dataset cover?
Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Tamil, Telugu and Urdu (native and romanised), plus Assamese, Bodo, Dogri, Kashmiri, Konkani, Maithili, Manipuri, Sanskrit, Santali and Sindhi (native script).
Is the dataset real Indian-language text?
No. All of it is machine-translated from English, with an AI model scoring each translation. The authors treat natural Indian-language text as a complement.
Can I download the Indic dataset?
The dataset is described as “releasing soon” and no licence has been named yet. The open-source pipeline used to generate it is available on GitHub.
Did the dataset improve AI performance?
NVIDIA reports gains of 4.98 to 9.82 percentage points on five Indian-language benchmarks after adding 10 billion Indic tokens to Nemotron 3.5 Super VL, with a small drop on one instruction-following test.