Synthetic translated data improves language models across twelve Indian languages

The authors release a 2.83-trillion-token corpus and multilingual evaluation suite, then show that combining synthetic and natural Indic text improves knowledge and reasoning over a public-data baseline.