---
type: "post"
id: "scaling-indic-pretraining-2-8-trillion-tokens-and-cc2843"
slug: "scaling-indic-pretraining-2-8-trillion-tokens-and-cc2843"
title: "Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline"
description: "Thesis: The announcement makes the generation process part of the data release: its value is not only a large Indic corpus, but a reproducible workflow for extending and auditing it."
retrieval_nugget: "Thesis: The announcement makes the generation process part of the data release: its value is not only a large Indic corpus, but a reproducible workflow for extending and auditing it."
published_at: "2026-10-16"
updated_at: "2026-10-16"
record_date: "2026-10-16"
date_kind: "published_at"
topics: ["models","open-models","open-source"]
source_urls: ["https://huggingface.co/blog/nvidia/scaling-indic-pretraining"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:scaling-indic-pretraining-2-8-trillion-tokens-and-cc2843"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:d85f8e7bf1492ffe2d29d42a741d8b12f0d0c0032051c8cb14b7c17a9c4e3e72","reviewed_at":"2026-10-16T03:00:13.440Z","source_evidence_count":1,"verified_claim_count":2,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":2,"implication_count":1,"watch_condition_count":1,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"The announcement makes the generation process part of the data release: its value is not only a large Indic corpus, but a reproducible workflow for extending and auditing it.","observed_facts":[{"text":"The source describes approximately 2.8 trillion tokens across 22 Indian languages, including 2,601.0 billion native-script tokens and 224.6 billion romanized tokens.","source_urls":["https://huggingface.co/blog/nvidia/scaling-indic-pretraining"]},{"text":"The source says records preserve source identifiers and describes translation, quality assessment, and targeted repair in the data-generation workflow.","source_urls":["https://huggingface.co/blog/nvidia/scaling-indic-pretraining"]}],"mechanism":"That workflow turns curated English source material into synthetic Indic text while retaining a traceable link to the source and adding quality gates.","why_now":"The publication pairs a sizeable corpus with an open recipe that teams can adapt to their own source material, broadening the practical unit of contribution from a fixed dataset to a repeatable data-generation process.","implications":["Model builders can evaluate the corpus as one input to Indic pretraining and adapt the pipeline, rather than treating the published token count as a ready-made training result."],"evidence_boundary":"Those results compare checkpoints across a stage combining synthetic and natural Indic data; they do not isolate the released corpus's causal contribution, and the token counts depend on the stated tokenizer.","watch_conditions":["Watch for controlled evaluations that separate corpus effects from natural-data contributions and test the full released dataset across languages and scripts."],"related_records":[],"new_branch_reason":"No prior New Runtime record is supplied for this specific corpus-and-generation-pipeline announcement; the distinct record tracks its scale, reproducible workflow and limits of the reported evaluation."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/scaling-indic-pretraining-2-8-trillion-tokens-and-cc2843/","markdown":"https://newruntime.com/posts/scaling-indic-pretraining-2-8-trillion-tokens-and-cc2843.md","json":"https://newruntime.com/posts/scaling-indic-pretraining-2-8-trillion-tokens-and-cc2843.json"}
---

# Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline

## Retrieval answer

Thesis: The announcement makes the generation process part of the data release: its value is not only a large Indic corpus, but a reproducible workflow for extending and auditing it.

Thesis: The announcement makes the generation process part of the data release: its value is not only a large Indic corpus, but a reproducible workflow for extending and auditing it.

Observed facts: NVIDIA and AI4Bharat describe approximately 2.8 trillion tokens spanning 22 Indian languages, including 2,601.0 billion native-script tokens and 224.6 billion romanized tokens. They say records preserve source identifiers and that the pipeline translates documents, assesses output quality, and repairs low-scoring translations. Mechanism: That workflow turns curated English source material into synthetic Indic text while retaining a traceable link to the source and adding quality gates. The authors report a 75/100 acceptance threshold for one twelve-language workflow; this is a rubric score, not a translation-accuracy percentage.

Why now: The publication pairs a sizeable corpus with an open recipe that teams can adapt to their own source material, broadening the practical unit of contribution from a fixed dataset to a repeatable data-generation process. First implication: Model builders can evaluate the corpus as one input to Indic pretraining and adapt the pipeline, rather than treating the published token count as a ready-made training result.

The source reports that 10 billion Indic tokens from a mixture of synthetic and natural text were used in a continued-pretraining stage for Nemotron 3.5 Super VL, with scores rising on five reported Indic benchmarks. Evidence boundary: Those results compare checkpoints across a stage combining synthetic and natural Indic data; they do not isolate the released corpus's causal contribution, and the token counts depend on the stated tokenizer. First watch condition: Watch for controlled evaluations that separate corpus effects from natural-data contributions and test the full released dataset across languages and scripts.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- topic hub: [Open Models](https://newruntime.com/topics/open-models.md): Explore the open-models topic hub.
- topic hub: [Open Source](https://newruntime.com/topics/open-source.md): Explore the open-source topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)