NVIDIA Introduces Synthetic Corpus for 22 Indian Languages
NVIDIA, IIT Madras and AI4Bharat announced a 2.83-trillion-token synthetic corpus for Indian languages. The full dataset is not yet publicly available, but the company has released the process used to generate it.
Nemotron Indic corpus is a new synthetic dataset that NVIDIA introduced on October 16 in collaboration with IIT Madras and its AI4Bharat laboratory. The project, designated Nemotron-Pretraining-Indic-v1, contains approximately 2.83 trillion synthetic tokens for 22 Indian languages and 34 language-script configurations. NVIDIA also released the tooling workflow used to create the data. However, the corpus itself is not yet publicly available, according to the company.
The data was created from curated English sources translated into Indian languages in their original scripts. For twelve languages, romanized variants were also added, meaning text written in the Latin alphabet. The goal is to expand the material available for pretraining language models for languages that have substantially less naturally produced digital text than English.
Nemotron Indic corpus and the open generation workflow
NVIDIA published a source project called the Nemotron Bharat translation plugin for the NeMo Data Designer tool. The workflow covers translating source documents, automated quality assessment, translation corrections and, for selected languages, transliteration into the Latin alphabet.
According to the documentation, traceability of the source documents is an important part of the process. This could make it easier for teams working with multilingual data to repeat or audit the workflow. The published plugin is available in the NVIDIA-NeMo/DataDesignerPlugins repository, where NVIDIA links to the tool used for this project.
However, the announcement does not mean that the full 2.83-trillion-token dataset is already available. NVIDIA says it will release the corpus later. The company has not yet confirmed a specific release date, licensing terms, distribution method or the size of individual files.
Results on the Nemotron 3.5 Super model
NVIDIA’s technical report describes using 10 billion tokens from a mixture of synthetic and natural Indian data for continued pretraining of the Nemotron 3.5 Super model.
On five benchmarks, NVIDIA reported improvements ranging from 4.98 to 9.82 percentage points. These are results measured by the company, and independent verification of the quality of the full corpus or reproduction of the reported benchmark improvements has not yet been located.
Translations do not replace natural content
A large volume of synthetic data may be particularly useful for languages with little available training text. However, synthetic translation is not the same as content created directly by users of a given language and in a given environment. The project’s authors themselves state that automated checks and sampled human reviews do not rule out translation errors.
According to the authors, broader validation by native speakers will be needed for ten languages. This is also relevant when evaluating the outputs of future models, since quality may vary between languages, scripts and types of text.
NVIDIA plans to release the Nemotron-Pretraining-Indic-v1 dataset later. Licensing terms, a dataset card, independent evaluations of individual languages and scripts, and the practical usability of the released plugin in NeMo Data Designer will also be important.
Sources
- Hugging Face Blog / NVIDIA – Announcement of the corpus, collaboration with AI4Bharat and IIT Madras, description of the pipeline, and notice that corpus release is still planned.
- NVIDIA Nemotron 3.5 Super and Lightning Technical Report – Token count, language coverage, training mixture and manufacturer-reported benchmark results after continued pretraining.
- NVIDIA Nemotron documentation — Stage 0: Pretraining – Technical description of the four-step workflow, corpus scope and the status of the planned release of the data and pipeline.
- NVIDIA-NeMo/DataDesignerPlugins – Evidence that the NVIDIA-managed plugin repository for NeMo Data Designer exists and that NVIDIA links to the plugin used in this work.
Verified and updated: October 16, 2026 15:24
