Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline

What is in the corpus? From English source documents to Indic training data The twelve-language workflow The ten-language workflow What changed after Indic continued pretraining?