Start Customizing NVIDIA Nemotron 3 Nano with Prime Intellect Lab in Minutes

Model customization typically faces multiple challenges such as infrastructure, technical expertise, and domain knowledge. The collaboration between NVIDIA and Prime Intellect Lab aims to address this: through Prime Intellect Lab's training-as-a-service platform, developers can complete local setup in about five minutes and launch a hosted reinforcement learning customization workflow.

This workflow does not require managing local GPU clusters; Prime Intellect Lab handles hosted deployment, training, and inference infrastructure. Developers only need a development environment (curl and Python 3), a Prime Intellect account, and an internet connection to get started.

The tutorial uses a Python math task as an example, demonstrating the standard customization workflow: first obtain baseline performance, then train a LoRA adapter using reinforcement learning with verifiable rewards (RLVR), and finally re-evaluate the model.

Baseline evaluation uses 32 examples with one rollout per example, showing the model initially performs poorly. After training, the model's accuracy improves significantly, but the tutorial explicitly notes that this is only an example; actual results depend on the specific task and evaluation method.

This workflow also applies to the Nemotron 3 Super and Ultra models. The tutorial also provides a companion notebook for a more complex coding task.

The NVIDIA Nemotron 3 series is released as open model resources, including weights, data, and training recipes. Developers can access NVIDIA Nemotron developer resources, read "Inside NVIDIA Nemotron 3," and check the Nemotron repository.

These resources provide a detailed view of model construction and post-training, including the data types and environments used, helping developers understand how their customization should differ from the base model. The open ecosystem supports reproducibility, transparency, and adaptation for specific use cases.

Model customization has multiple methods, including supervised fine-tuning, parameter-efficient fine-tuning, preference optimization, reinforcement learning, or a combination of techniques. This tutorial focuses on reinforcement learning with verifiable rewards (RLVR), using Python tools (numpy, sympy, scipy) for computation in a Python math environment.

The environment defaults to 100 rounds, but the tutorial limits each rollout to 5 assistant turns to penalize repeated tool calls without producing a final response. This design choice affects training behavior and final model performance.