Nemotron-RL-Super: Training Blends for Nemotron-3-Super-120B-A12B

NVIDIA released this collection of dataset blends in March 2026 to document the specific data mixtures used for Reinforcement Learning (RL) training of the Nemotron-3-Super-120B-A12B model. The data is organized into six distinct training stages including Reinforcement Learning from Verifiable Rewards (RLVR), Software Engineering (SWE), and Reinforcement Learning from Human Feedback (RLHF).

Use Cases

• Replicating RL training stages for large language models using the specified mixing ratios
• Analyzing the composition of SWE 1 and SWE 2 stages for software engineering tasks
• Evaluating the impact of RLVR (Verifiable Rewards) data on model performance

Strengths

• Provides exact mixing ratios for six distinct RL training stages
• Covers specialized domains including Software Engineering (SWE) and Verifiable Rewards (RLVR)
• Released by NVIDIA to provide transparency for Nemotron-3-Super-120B-A12B training

Limitations

• Missing raw row counts and specific column schemas in the provided metadata
• Excludes 'additional data' mentioned in the source description
• High dependency on external component datasets for full replication

Provenance

Source

NVIDIA

Collection Method

Curated blends of existing datasets

Freshness

Last updated March 2026.

Geography

United States