Nemotron Post-Training Dataset v2

Nemotron Post-Training Dataset v2 is NVIDIA's 2025 post-training collection for supervised fine-tuning and reinforcement learning, released on Hugging Face in August 2025. It gathers prompts and synthetically generated responses across mathematics, code, STEM, general reasoning, and instruction following, then expands the instruction-following data into five additional languages (German, Italian, Spanish, French, and Japanese) alongside the English splits. Responses were produced by multiple models, including DeepSeek-R1-0528 and Qwen variants, so the set reflects distilled reasoning traces rather than human-written answers.

Access is through the Hugging Face dataset page, where the data is published as Parquet and organised into named splits (math, code, stem, chat, and a multilingual split per target language). Reported split sizes include roughly 239,000 mathematics rows, 175,000 code rows, 355,000 STEM rows, and 627,000 chat rows, with each multilingual split adding close to a million rows. You filter by split or metadata and download in bulk; there is no query API.

For RAG and AI work this is a fine-tuning and alignment resource, not a retrieval corpus. Use it to instruction-tune or post-train an assistant, to add multilingual reasoning coverage, or as a pool of reasoning traces for distillation. It pairs naturally with open base models and complements the pretraining corpora in this directory.

On licensing, most of the dataset is CC-BY-4.0, so commercial use is permitted provided you attribute. Two caveats matter: a WildChat-derived subset is under ODC-BY, and a StackOverflow-derived subset is under CC-BY-SA, which carries a share-alike obligation on that slice. Check the per-record licence metadata before redistributing derived data. This is distinct from our Nemotron-CC-v2 listing, a gated pretraining web corpus under NVIDIA's data agreement; this post-training set is openly licensed and aimed at the fine-tuning stage.