An arXiv announcement introduces Nemotron-Pretraining-Indic-v1, a synthetic pretraining corpus of approximately 2.83 trillion tokens. The dataset is designed to cover all 22 scheduled languages of India, addressing the need for scalable multilingual data in building capable language models for the region.
The abstract frames the corpus as part of a broader effort to improve both pretraining and evaluation for Indic LLMs. By providing a large synthetic resource, the authors aim to reduce the bottleneck of curated multilingual data and support more robust benchmarking across Indian languages.
As the announcement is brief, details on corpus construction, quality filtering, and evaluation methodology are not yet available. The stated scale and language coverage, however, signal a significant investment in Indic-language pretraining resources.