Open datasets for moving an agent from SFT to reinforcement learning

Use verifiable rewards where you can and a reward model elsewhere

NVIDIA publishes agentic RL datasets built for this step. Nemotron-RL-Agentic-Function-Calling-Pivot-v1 covers function calling, Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1 covers tool use inside conversations, Nemotron-RL-Agentic-SWE-Pivot-v1 covers software engineering, and Nemotron-RL-Agentic-Terminal-Pivot-v1 covers terminal tasks. The newer Nemotron-RL-Agentic-Conversational-Tool-Use-Seed-v1 adds seed policies, tool definitions, and scenarios for simulated conversations in hospitality, e-commerce, banking, and insurance. Supervised fine-tuning teaches an agent by imitation. RL scores what the agent actually does across a multi-step rollout, so the data has to come with a way to grade the outcome.

When a task has a checkable outcome, such as a passing test or a correct database state, score it with a verifier at rollout time. NVIDIA NeMo Gym is an open-source library of environments and verifiers for this, and NVIDIA NeMo RL runs the training. When outcomes cannot be checked automatically, train a reward model on preference data and let it score trajectories. Nemotron-RLHF-GenRM-v1 supports generative reward models, which explain a judgment along with the score. Many teams combine the two: verifiable rewards where they can, and a learned reward model for the rest. Start with the RL dataset closest to your agent's production workload.

Datasets to start with

• Nemotron Reinforcement Learning collection: NVIDIA's RL datasets with verifiable rewards across math, code, agentic, and instruction tasks.

• Nemotron-RL-Agentic-Function-Calling-Pivot-v1: RL data for function calling.

• Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1: RL data for tool use inside conversations.

• Nemotron-RLHF-GenRM-v1: data for generative reward models, for outcomes a verifier cannot check.

Key takeaways

• RL needs a reward for each rollout. SFT demonstrations alone do not provide one.

• NVIDIA publishes agentic RL datasets for function calling, conversational tool use, software engineering, and terminal work.

• NVIDIA NeMo Gym supplies open environments and verifiers that score rollouts.

• Where outcomes cannot be checked, train a reward model on preference data such as Nemotron-RLHF-GenRM-v1.

• Start with the RL dataset closest to your production workload.