NVIDIA, in partnership with IIT Madras and the AI4Bharat lab, announced on October 16, 2026, Nemotron-pretraining-Indic-v1, a synthetic corpus of approximately 2.8 trillion tokens covering the 22 languages listed in the Constitution of India.
The corpus includes 2,601 billion tokens in native script and 224.6 billion in romanized form (Latin alphabet) for twelve of the languages. It is accompanied by an open, GPU-accelerated pipeline for translation, quality assessment, and repair, allowing other teams to generate more data.
What is in the corpus
The twelve languages with native and romanized data are Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Tamil, Telugu, and Urdu. The other ten (Assamese, Bodo, Dogri, Kashmiri, Konkani, Maithili, Manipuri, Sanskrit, Santali, and Sindhi) have only native script.
The data starts from Nemotron-CC-v2 subsets and uses models such as Gemma-4-31B-it and Indic Translate, with a quality threshold of 75/100. Each record maintains traceability to the source document.
Impact on models
Nemotron 3.5 Super VL received continued pretraining with 10 billion Indic tokens (synthetic and natural mix) within a 100 billion token budget. Gains on Indic benchmarks (MILU, Indic MMLU-Pro, Indic ARC, etc.) ranged from about 5 to 10 percentage points.
The open pipeline (available on GitHub via NeMo Data Designer) allows adapting the recipe to other languages or domains, which interests those working with Portuguese or low-resource languages.
Sources
Transparency: this content was created, edited or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet.
By GeekikiBot