Nemotron-RL-Science-v1: Reinforcement Learning Problems for Science Reasoning

Nemotron-RL-Science-v1 provides a reinforcement learning dataset for training models on science reasoning tasks. Each example includes a problem, a reference answer, and a verifiable RL environment configuration for training with verifiable rewards. The dataset was created by NVIDIA and covers three science domains: Physics, Biology, and Chemistry.

Use Cases

• Training a policy model with verifiable rewards based on the provided RL environment configurations.
• Benchmarking reinforcement learning agents on science reasoning problems across physics, biology, and chemistry domains.
• Developing answer-extraction methods using the provided templates for structured output.
• Studying agent/verifier interactions in a controlled, reward-driven setup for scientific problem-solving.

Strengths

• Includes verifiable RL environment configurations (agent prompt, verifier reference, answer-extraction template) for each problem.
• Covers three distinct science domains: Physics, Biology, and Chemistry.
• Provides both problems and reference answers for supervised or reward-based learning.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

NVIDIA

Freshness

Last updated 2026-06-04 06:46:39; freshness should be verified.