What is announced
NVIDIA and AI4Bharat at IIT Madras describe roughly 2.8 trillion tokens translated from English for 22 Indian languages. All 22 have native-script translations; twelve also have Latin-script variants. Native-script describes the writing system, not naturally authored local text. Variants can share one source document; tokens are not a count of unique originals.
Check release status and quality evidence
The October 16 (UTC) post says the corpus is “releasing soon.” Its linked GitHub plugin page and documentation returned 404 during our check without signing in, leaving public pipeline access unverified.
A twelve-language workflow uses the same model to translate and judge. Its 75/100 filter is a rubric score, not accuracy. Reported pretraining gains combine synthetic and natural data; they do not isolate this corpus’s effect.
Our take
Editorial analysis: review useful examples before scaling a language-data project. Ask fluent reviewers familiar with the intended audience to check meaning, local context, and script conventions. Keep variants from one source on the same side of a training/evaluation split. Check release terms and access before budgeting work. We have not downloaded the data, run the pipeline, or tested the reported gains.