Nemotron-RL-Math-v2: Curated Mathematical Problems for Reinforcement Learning

Nemotron-RL-Math-v2 is a small curated set of mathematical problems selected for reinforcement learning workflows. The dataset is designed for Reinforcement Learning from Verifiable Rewards (RLVR) and was created by NVIDIA. Problems are sourced from AoPS and StackExchange-derived math data held out from the Nemotron-SFT-Math-v4 SFT set.

Use Cases

• Training RL agents on mathematical problem-solving based on the described verifiable answers.
• Benchmarking RLVR methods based on the curated set of problems with validation signals.
• Fine-tuning language models for math using a reinforcement learning reward signal derived from the dataset's structure.

Strengths

• Designed for RL workflows with verifiable answers, a specific training paradigm.
• Curated from established sources like AoPS and StackExchange.
• Created by NVIDIA, suggesting institutional backing.

Limitations

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

Provenance

Source

NVIDIA, with problems sourced from AoPS and StackExchange.

Collection Method

Curated selection from held-out math data.

Freshness

Last updated 2026-06-04 04:54:18; freshness should be verified.