NVIDIA has announced Nemotron-pretraining-Indic-v1, a synthetic text corpus it says contains roughly 2.8 trillion tokens across 22 Indian languages, alongside a public data-generation workflow. Developed with IIT Madras and AI4Bharat, the project was announced on October 16, 2026, and covers 34 language-script configurations. 1 2
The announcement establishes the reported scale and documents how the data was generated, but full corpus availability remains unresolved. At the October 18 research cutoff, AI4Bharat described the data as already open, while NVIDIA’s training documentation still designated it as announced for release. The supplied research did not verify a definitive corpus download. 3 4
What the 2.8-trillion-token figure measures
NVIDIA reports an exact inventory of 2,825.6 billion retained output tokens: 2,601.0 billion in native scripts and 224.6 billion in romanized form, counted using the Gemma-4 tokenizer. Its technical report distinguishes that inventory from both input-processing volume and the number of tokens used to train the Super model. The 438.4 billion tokens recovered through translation repair are already included—not an additional contribution to the total. 1 5
According to the announcement, generation starts with English documents from 15 Nemotron-CC-v2 subsets, including web text, question-and-answer material, mathematics and code. Source identifiers are retained. Native-script output therefore does not mean the documents were originally authored in an Indian language. Because translations and script variants can share source documents, the headline count should not be read as an equivalent volume of distinct knowledge. 1
The reported inventory exceeds earlier resources such as AI4Bharat’s approximately 251.3-billion-token Sangraha and Krutrim’s 540-billion-token BhashaKritika. But those are not like-for-like comparisons: Sangraha combines multiple kinds of material and includes English, while BhashaKritika uses several synthetic-generation methods and says only part of its data is publicly shared. Token volume alone does not establish better quality, diversity or training efficiency. 5 6 7
Two translation pathways, with different coverage
NVIDIA describes a twelve-language pathway in which Gemma-4-31B-it translates documents and checks their quality, with rewriting for low-scoring translations. This group includes both native-script and romanized material. The corpus acceptance threshold is at least 75 out of 100 on a quality rubric—not a measured 75% accuracy rate. 4 1
For another ten languages, NVIDIA’s engineering documentation describes using Bodhan AI’s Indic Translate where Gemma’s translation quality fell short. This group receives native-script data only. The developers report substantial degradation for Santali documents longer than approximately 2,048 tokens and describe translating Manipuri through Bengali script before deterministic conversion to Meitei Mayek. These are developer-reported observations, not independently reproduced findings. 8 4
The public plugin reference exposes configurable translation, evaluation and rewrite prompts, separately selectable translator and evaluator models, and intermediate-language stages. It also supports translating overlapping document segments and evaluating the reconstructed result. That windowing mode disables rewriting; separately, rewrite attempts default to zero. Installing the plugin therefore does not automatically reproduce every production setting. 9
The documented contribution is a reusable generation and assessment workflow. It is not, by itself, proof of identical output quality or affordable regeneration at trillion-token scale. Using the same model to translate and judge also raises a potential risk of correlated errors. Configurable evaluators permit alternative checks, but do not establish that independent judging was used throughout production. 9 4
Reported gains came from a smaller, mixed training stage
NVIDIA says its continued-pretraining stage used 100 billion tokens, including 10 billion synthetic-plus-natural Indic tokens. The other 90 billion maintained the preceding training mixture’s proportions. The stage did not consume the entire announced 2.8-trillion-token corpus. 4
The company reports the following before-and-after checkpoint results in its official pretraining documentation. These are author-reported comparisons, not independent benchmark reproductions. 4
Benchmark Languages evaluated Before After Change MILU 10 72.48 77.90 +5.42 points Indic MMLU-Pro 12 51.83 56.81 +4.98 points Indic ARC-Challenge 12 84.17 92.83 +8.66 points Indic ARC-Easy 12 90.73 96.77 +6.04 points Indic GSM8K, strict 11 72.67 82.49 +9.82 pointsThe comparisons support the combined continued-pretraining intervention, not a causal claim that synthetic data alone produced the gains. Natural Indic material and substantial additional training were also part of the intervention. In a separate final-model comparison, NVIDIA’s announcement reports a 2.04-point decline on IFEval 1.5 versus Nemotron 3 Super, so the broader results are not uniformly positive. 4 1
Evaluation coverage also falls short of the corpus’s full scope. The announcement describes controlled studies of first-pass accepted data from the twelve-language pathway, excluding recovered rewrites and the additional ten languages. Sampled human-review coverage varies by language, and source-ID deduplication addresses processing duplicates rather than semantic duplication. Translation at this scale does not replace locally authored knowledge. 1
AI4Bharat and NVIDIA are also introducing Indic-LLM-Eval-Suite, combining benchmark datasets, evaluation implementations and construction methodology. Its benchmark-translation process includes automated assessment and human review. That provides evaluation infrastructure, but it is a companion project from the same collaborators—not independent validation of the corpus. 3
Access, licensing and reproducibility remain open
Beyond the conflicting release descriptions, the dataset-specific license has not been verified. The public companion-report repository displays CC BY 4.0, but that designation does not establish the license governing the announced corpus. 10
Nor does a public generation workflow make the complete model-training program reproducible. NVIDIA states that full Super 3.5 pretraining depends on nonpublic teacher logits and training mixtures. 4
The supplied research identifies several questions still requiring answers: Where are the definitive corpus files, manifest and license? Will production prompts, model revisions and filtering settings be version-pinned? What are the per-language human error rates and human–judge agreement rates? Fixed-budget experiments are also needed to separate synthetic from natural data, repaired from first-pass output, and romanized from native-script contributions. Other unresolved issues include cross-language duplication, benchmark-contamination controls, evidence of gains for the additional ten languages, and generation costs per accepted token.
For now, the evidence establishes an announced trillion-token-scale corpus, an inspectable generation workflow and NVIDIA-reported training improvements—not verified full-dataset access, independent replication or a measured synthetic-data-only effect.
Research boundary: This draft uses only the supplied research package. That research did not execute the pipeline, download the full corpus or independently reproduce benchmarks. The companion Indic report’s file listing was accessible, but its PDF download failed, so detailed ablations in that report were not directly inspected.