NVIDIA and AI4Bharat announce Nemotron-pretraining-Indic-v1, a ~2.8T-token synthetic corpus across 22 Indian languages, with open generation pipeline to follow

NVIDIA and IIT Madras' AI4Bharat announced Nemotron-pretraining-Indic-v1, a synthetic corpus of about 2.8 trillion tokens (2,601.0B in native script, 224.6B romanised) covering 22 Indian languages in 34 language-script configurations. It was made by machine-translating 15 English Nemotron-CC-v2 subsets with Gemma-4-31B-it via NeMo Data Designer, then scoring quality (threshold 75/100) and repairing weak output. Neither the dataset nor the generation pipeline (prompts, workflows, quality checks) has been released. Both are described as "releasing soon". NVIDIA says it has already used 10B tokens of the data for continued pretraining of Nemotron 3.5 Super VL.

If released as described, this would give researchers and developers a large, openly available Indic pretraining resource, which matters because data scarcity is a real constraint for Indian language models. The open pipeline could also let others build more data for their own languages or domains. Today, though, nobody outside NVIDIA and AI4Bharat can use or inspect any of it, so there are no decisions to make yet. Because the text is translated from English, it may carry English-centric content and translation artefacts rather than native Indian usage, and the effect on model quality is unproven.

The "open" framing is premature, since the data and recipe are still promised rather than delivered. The headline token count measures volume, not quality or cultural coverage, and the 75/100 quality score comes from the builders' own automated checks. The only usage evidence is 10B tokens in continued pretraining of one NVIDIA model, with no benchmark results shown here.

1. 01Whether the dataset and pipeline actually appear on Hugging Face and GitHub, and when, along with the licence terms and whether the full 2.8T tokens are included.
2. 02Independent evaluations on Indic benchmarks, comparing models trained on this data against baselines, including results for low-resource languages and native speaker reviews of fluency and cultural accuracy.
3. 03Published results showing how much the 10B-token continued pretraining improved Nemotron 3.5 Super VL on Indic tasks, and whether other groups reproduce gains using the released pipeline.