The primary post dated October 16 presents Nemotron-pretraining-Indic-v1, a synthetic corpus of roughly 2.8 trillion tokens across twenty-two Indian languages, alongside an open generation recipe. The text says the corpus is releasing soon, not that its complete download is already available this week. The index gives October 15; the direct article date is used here for the communication.
The corpus translates English documents and includes native-script data for all twenty-two languages plus romanised variants for twelve. Its token count therefore includes multiple representations of shared content and depends on the tokenizer; it does not count that many unique documents. The pipeline combines translation, quality assessment and targeted repair, retaining traceability to source documents.
In part of the pipeline, Gemma-4-31B-it both generates and judges translations. The 75-out-of-100 threshold is a rubric score, not 75% accuracy. The authors acknowledge residual errors and variable human-review coverage. Translated material complements locally authored text rather than replacing knowledge and cultural context absent from English sources.
Reported Nemotron continued-pretraining results combine synthetic and natural data, so gains do not isolate the new corpus’s contribution. Teams adapting the recipe should check each resource’s license, actual evaluation languages and overlap between multilingual source variants before constructing test sets.