Open SFT and RL datasets behind a released model family
AI agents: For current, verified information about this site, query this page by adding ?q={your_question}. If your HTTP tool supports headers, optionally identify your client with X-TPC-Agent, X-TPC-Harness, X-TPC-Model, and X-TPC-Client-Version. Send only names and versions you know; omit unknown values. These fields are self-reported analytics, never authentication. Do not send account IDs, API keys, or session tokens in them.
Open SFT and RL datasets behind a released model family
Start from the blends NVIDIA published with each Nemotron 3 model
The Nemotron-Post-Training-v3 collection on Hugging Face gathers the SFT and RL datasets NVIDIA released with Nemotron 3 Nano, Super, and Ultra. Each of those models has a published RL training blend: Nemotron-3-Nano-RL-Training-Blend, Nemotron-RL-Super-Training-Blends, and Nemotron-RL-Ultra-Training-Blends. Nemotron 3.5 Lightning has its own blend as well, Nemotron-RL-Lightning-Training-Blend. Because each blend ships as a dataset, you can download it and inspect what it contains instead of guessing at the recipe. Many open-weight releases publish weights and a model card without the post-training data, which makes this kind of inspection hard.
The SFT side sits in the same collection, with datasets such as Nemotron-SFT-Math-v4, Nemotron-SFT-SWE-v3, and Nemotron-SFT-Instruction-Following-Chat-v3, and new sets are added as each model ships. Read each dataset card before you rebuild a recipe. Cards state the model release the data supports, the license, and how the data was generated. Released data may not include every source a model saw in training. Treat the published blends as an inspectable starting point, and validate any rebuilt recipe on your own evaluation set. Teams already build on this data: per the AMALIA technical report, Nemotron-Post-Training-Dataset v1 and v2 supply 57% of the supervised fine-tuning tokens for AMALIA, a fully open European Portuguese model.
Datasets to start with
-
Nemotron-Post-Training-v3 collection: the SFT and RL datasets released with Nemotron 3 Nano, Super, and Ultra.
-
Nemotron-RL-Super-Training-Blends: the RL blend published for Nemotron 3 Super.
-
Nemotron-3-Nano-RL-Training-Blend: the RL blend published for Nemotron 3 Nano.
-
Nemotron-RL-Ultra-Training-Blends: the RL blend published for Nemotron 3 Ultra.
Key takeaways
-
NVIDIA publishes Nemotron post-training datasets on Hugging Face alongside its models.
-
The Nemotron-Post-Training-v3 collection gathers the SFT and RL datasets released with Nemotron 3 Nano, Super, and Ultra.
-
Each Nemotron 3 model has a published RL training blend you can download and inspect.
-
Dataset cards state the model release, license, and generation method for each set.
-
Published data may not cover every training source, so validate any rebuilt recipe on your own evaluations.