Nemotron RL Ultra Training Blends: Reinforcement Learning Data for Post-Training

NVIDIA's Nemotron-3-Ultra post-training recipe uses these Reinforcement Learning and Multi-teacher On-Policy Distillation training-data blends. Each prompt is paired with an agent or environment that returns a verifiable or judge-based reward, as consumed by the NeMo Gym agent framework. The dataset was last updated on June 4, 2026.

Use Cases

• Training or fine-tuning large language models via reinforcement learning based on the described reward mechanisms.
• Implementing Multi-teacher On-Policy Distillation (MOPD) using the provided training-data blends.
• Benchmarking RL training frameworks like NeMo Gym with the described prompt-reward pairings.
• Studying the effect of different reward signals (verifiable vs. judge-based) on model alignment.

Strengths

• Created by NVIDIA, a leading AI research institution, for a public model recipe.
• Designed for a specific, documented training framework (NeMo RL recipes and NeMo Gym).
• Last updated on June 4, 2026, suggesting recent maintenance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file size are unknown, which may limit suitability assessment.
• The description is partial, referencing an external page for full details.

Provenance

Source

nvidia

Collection Method

Likely generated as part of the Nemotron-3-Ultra model development and post-training process.

Freshness

Last updated 2026-06-04 11:34:41