---
title: "AI4Bharat and NVIDIA Release 2.8 Trillion-Token Indic Pretraining Corpus"
slug: "ai4bharat-and-nvidia-release-2-8-trillion-token-indic-pretraining-corp"
date: "2026-10-16"
category: "AI & Automation"
author: "Marcus «Glitch» Reed"
correspondent: "Marcus «Glitch» Reed"
correspondentRole: "Open Source & Hacker Scout"
readingTimeMinutes: 1
tags: ["AI","News"]
originalSourceUrl: "https://huggingface.co/blog/nvidia/scaling-indic-pretraining"
canonicalUrl: "https://aurav.ae/ai4bharat-and-nvidia-release-2-8-trillion-token-indic-pretraining-corp"
publisher: "Aura-V Intelligence (aurav.ae)"
---

# AI4Bharat and NVIDIA Release 2.8 Trillion-Token Indic Pretraining Corpus

> AI4Bharat and NVIDIA have open-sourced a 2.8 trillion-token dataset covering 22 Indian languages, together with a GPU-accelerated generation pipeline.

## Key Takeaways (TL;DR)

- Nemotron-pretraining-Indic-v1 provides 2.8 trillion tokens across 22 Constitution-listed Indian languages.
- The release includes a reusable GPU-accelerated pipeline for translation, quality assessment, and script conversion.
- Native-script data totals 2,601 billion tokens; romanized text adds 224.6 billion tokens.

AI4Bharat and NVIDIA announced the open-source release of **Nemotron-pretraining-Indic-v1**, a corpus of approximately **2.8 trillion tokens** spanning the 22 languages listed in India’s Constitution, representing the first language for 97 % of the population (about 1.2 billion people) [Hugging Face Engineering Blog](https://huggingface.co/blog/nvidia/scaling-indic-pretraining). The dataset is accompanied by a reusable GPU-accelerated pipeline that handles document-level translation, quality assessment, and targeted repair, allowing developers to generate additional synthetic data.

The corpus contains **2,601.0 billion native-script tokens** and **224.6 billion romanized tokens**, covering twelve languages in both scripts and ten languages in native script only. Workflows employ NeMo Data Designer and Gemma-4-31B-it for translation and scoring, retaining outputs that meet a 75-point quality threshold. The pipeline preserves technical content such as equations, tables, and code, and each synthetic record is traceable to its source document.

Our read: The joint effort supplies a massive, verifiable data source for Indic language models, lowering the barrier for B2B firms to localize AI services across India’s diverse linguistic landscape.

---

*Original Think-Tank Source: [Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline](https://huggingface.co/blog/nvidia/scaling-indic-pretraining)*