Nemotron RL Reasoning Gym V1 with 15,000 Samples

15,000 samples span 104 reinforcement learning environments for tasks in algebra, arithmetic, computation, cognition, geometry, graph theory, logic, and common games. Generated procedurally using the Reasoning Gym library by NVIDIA, this dataset is designed for commercial use to improve reasoning capabilities.

Use Cases

• Train reinforcement learning agents on procedurally generated tasks across 104 environments like algebra and logic.
• Benchmark model reasoning performance on algorithmically verifiable tasks from domains including geometry and graph theory.
• Fine-tune language models for text generation on cognitive and computational problem-solving samples.
• Develop multi-task learning systems using the broad range of domains, including arithmetic and common games.

Strengths

• Contains 15,000 samples for training and evaluation.
• Covers a broad range of 104 distinct RL environments and domains.
• Procedurally generated with algorithmically verifiable tasks, ensuring consistency.

Limitations

• Unknown column structure and sample data format limits immediate usability.
• Lacks explicit information on dataset size, file formats, and licensing details.
• Potential for synthetic generation artifacts not grounded in real-world observations.

Provenance

Source

NVIDIA

Collection Method

Procedurally generated using the Reasoning Gym library.

Time Range

null

Freshness

Last updated March 2026.

Geography

null