Nemotron-3-Nano-RL: Multi-Domain Training Blend for 30B Models

NVIDIA released the Nemotron-3-Nano-RL-Training-Blend in late 2024 as a curated collection for training the Nemotron-3-Nano-30B-A3B model. The dataset aggregates five specialized sub-datasets with precise mixing ratios, including 20% multiple-choice knowledge and 17% instruction following. It is designed specifically for Reinforcement Learning (RL) workflows in large language model development.

Use Cases

• Fine-tuning models for instruction following using the 17% instruction_following subset
• Optimizing multiple-choice reasoning via the 20% knowledge-mcqa component
• Training agents for structured data generation using the 5% structured_outputs ratio

Strengths

• Defined mixing ratios for five distinct components (e.g., 0.20 for MCQA, 0.05 for structured outputs)
• Multi-domain coverage including competitive coding and workplace assistant tasks
• Official training source for the Nemotron-3-Nano-30B-A3B model

Limitations

• Lack of explicit column schema or sample data in the top-level metadata
• Potential domain bias toward NVIDIA-specific workplace assistant requirements

Provenance

Source

NVIDIA

Collection Method

Curated blend of existing NVIDIA-proprietary and open RL datasets

Freshness

Last updated in December 2025, reflecting current training methodologies for the Nemotron-3 model family.

Geography

United States