Nemotron Image Training v3 targets large-scale vision–language training workflows by bundling 76 focused subdatasets into a single, standardized release. Rather than supplying pixels inline, each subdataset ships a Megatron/Megatron-Energon–compatible conversation JSONL and a README describing upstream media sources, licensing, and the expected local layout—so the main deliverable is curated, model-ready multimodal supervision with explicit provenance and labels.
What Sets It Apart
- Broad, task-focused coverage: the collection spans QA, OCR, grounding, multi-image reasoning, chart/plot QA, and more, enabling unified training for diverse VLM capabilities rather than a single benchmark. This makes it easy to assemble mixed curricula (e.g., OCR-heavy + reasoning-heavy mixes) without reinventing formatting.
- Large but modular: ~6.9M rows and an estimated 39.56B tokens distributed across 76 subdatasets (approx. 9.25 TB referenced media). Each subdataset includes a dataset card and a conversation JSONL, so you can pick only the subsets you need and reproduce the original media layout.
- Labeling & verification transparency: data are a hybrid of human and synthetic sources, with model-labeling and verification passes noted per-subdataset (e.g., qwen/gemini/gpt labels), and clear license statements (mainly CC-BY-4.0 with a few GPL/BSD/CC BY-SA exceptions).
Who It’s For & Tradeoffs
Great fit if you need: curated multimodal supervision to pretrain or fine-tune VLMs at scale; modular subdatasets for targeted curricula; Megatron/Megatron-Energon–style JSONL ingestion. You get standardized conversational examples and provenance metadata, which reduces preprocessing friction. Look elsewhere if: you need datasets with embedded pixels (this release references upstream media only), if strict end-to-end commercial clearance is required for every sample (some subdatasets carry differing upstream licenses), or if you require uniformly human-verified labels—quality and label provenance vary by subset.
Practical Notes
- Owner/publisher: NVIDIA Corporation; first published 2026-04-28. License: primarily CC-BY-4.0, with specific subdataset exceptions (e.g., ChartQA labels under GPL-3.0; zhwiki includes CC BY-SA / GFDL constraints). Always consult each subdataset README before commercial use.
- Integration: JSONL format is compatible with Megatron-Energon loaders; media must be downloaded & arranged locally as described in each README so paths resolve at training time.
- Resource planning: referenced media volume is substantial (≈9249.07 GB); plan storage and data-access bandwidth accordingly.
Overall, v3 is a practical, modular package for teams building or scaling vision–language training pipelines where explicit media provenance, per-subdataset licensing, and standardized multimessage JSONL formatting matter more than immediate pixel availability.
