What happened
NVIDIA and AI4Bharat’s paper, “Scaling Synthetic Pre-training Data and Evaluation for Indic LLMs,” introduces Nemotron-Pretraining-Indic-v1 (NPI-v1) and the Indic-LLM-Eval-Suite. The paper describes NPI-v1 as an openly released synthetic pretraining corpus covering all 22 languages listed in the Eighth Schedule of the Indian Constitution. Its reported total is approximately 2.83 trillion tokens. The authors say the release is the largest openly released Indic pretraining corpus, a claim they qualify as being to the best of their knowledge.
The corpus contains 2.6 trillion tokens in native scripts and 225 billion romanized tokens, measured with the Gemma-4 tokenizer. The paper reports 2.44 billion documents across 34 language–script configurations. Twelve higher-resource languages have both native-script and Latin-script romanized configurations; the romanized material is made by transliterating a subset of accepted native-script translations and is about one tenth of their native-script volume. The other ten languages currently have native-script data only, with romanized versions planned for later releases. The paper uses BCP-47 language, script and region tags to identify configurations.
The authors describe a pipeline that translates curated English documents, evaluates translation quality, and repairs low-quality outputs while aiming to preserve technical content, document structure and traceability to source documents. The paper emphasizes that synthetic generation can introduce errors, unnatural language and structural damage, and presents automated checks and targeted repair as controls rather than assuming that fluent translated text is necessarily good training data. The data, generation recipes, evaluation resources and documentation are described as open source.
The companion Indic-LLM-Eval-Suite is an extensible multilingual benchmark collection covering chat, instruction following, reasoning, knowledge, code and translation. It combines existing Indic benchmarks with task-aware translations of English benchmarks. The paper says the benchmark pipeline applies automated quality assessment and targeted correction, followed by native-speaker review, and that the resulting suite is integrated into NVIDIA’s NeMo-Gym for standardized, reproducible evaluation.
To test data utility, the authors report continued-pretraining experiments across twelve languages, comparing corpus compositions under uniform training settings. They say mixing their synthetic corpus with natural Indic text improves knowledge and reasoning over a public-corpus baseline. They also study effects of data volume, script coverage and Indic training share on English performance. The paper characterizes the resulting checkpoints as experimental instruments for measuring data utility, not as state-of-the-art model releases.
What the archive adds
New Runtime’s earlier coverage on 2026-10-21 described the 2.8-trillion-token Indic pretraining effort and its open data-generation pipeline (Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline). This paper is a new, fuller research account of that known thread, adding a benchmark suite and reported continued-pretraining experiments. Earlier coverage of Nemotron 3.5 Super VL concerns a separate model release, not this corpus.
Trend signal
This does not yet fit a pattern New Runtime tracks. The release concerns multilingual training data and evaluation resources; none of the current pattern theses directly describes this development.
Why now
The authors frame the work as a response to the shortage of foundational data and comprehensive evaluation for Indian languages. They point to a large disparity between English resources and Indic-language text, and argue that translated material can scale coverage if its quality, structure and provenance are checked. The release pairs data construction with evaluation rather than treating corpus size alone as evidence of progress.
Why it matters
For teams building language technology for Indian users, the release offers a way to experiment with substantially more open training material and a broader evaluation set without first assembling every resource independently. The source-traceable workflow is also relevant beyond this particular corpus: multilingual data teams need to know what was translated, how it was checked, and where repairs were applied. The benchmark suite may help expose differences between languages and capabilities that a single aggregate score would hide.
The reported gains are promising but do not establish that synthetic data alone improves every model or product. The paper itself describes mixing synthetic and natural text, so teams should test the corpus against their own languages, tasks and baselines. The practical contribution is a set of resources for controlled experimentation—not a ready-made guarantee of better user-facing performance.
Limits and uncertainty
The results are claims made by the paper’s authors and have not been independently checked here. The reported corpus size uses the Gemma-4 tokenizer, so token totals should not be compared directly with counts produced by other tokenizers. The experiments described are continued pretraining across twelve languages, not evidence of uniform gains across all 22 languages or every downstream application. The authors call the corpus the largest open release to their knowledge; that is not an independently verified ranking. The paper also describes some romanized configurations as future work.
Skill to build now
If your team evaluates or trains for an Indic language, choose one target language and task, then compare a baseline using your current data with a controlled run that adds a small, traceable portion of synthetic translated text. Record script, data mix, tokenizer and evaluation set, and ask native speakers to review representative benchmark items. The paper’s central lesson is to test the quality and utility of translated data, not infer either from corpus scale.
Reader takeaway
NPI-v1 is a large open Indic-language corpus, but its significance is the combination of data, quality-control recipes and evaluation resources. Treat the reported improvements as a research result to reproduce for your own language and task, rather than as proof that a bigger synthetic corpus automatically yields a better product.