Fine-Tune Specialized LLMs with Unsloth

Verify GPU access

Confirm GPU access on your hardware platform before you continue with any fine-tuning workflow.

nvidia-smi

Expected output should show GPU information for your hardware platform.

Choose a fine-tuning method (intended)

Match the method to your goal and data size:

1. Parameter-efficient fine-tuning (LoRA / QLoRA) — small- to medium-sized datasets; lower VRAM and faster iteration
2. Full fine-tuning — large datasets when you need broader parameter updates for format, style, or guardrail behavior
3. Reinforcement learning — preference or reward-driven behavior change; see Unsloth's Reinforcement Learning Guide

This playbook does not ship hyperparameters, training scripts, or a dependency list for those runs.

Pick a starting model (intended)

When you are ready to train, choose an open model that fits your VRAM and task. Nemotron 3 Nano 30B-A3B is one option optimized for agentic fine-tuning workloads; download it from Hugging Face when you follow an upstream Unsloth or training guide.

Confirm license terms and any gated-model access before downloading.

Follow an upstream Unsloth workflow (intended)

When you move from this overview to hands-on training:

1. Follow Unsloth’s published guides and notebooks for your GPU generation
2. Keep other heavy GPU workloads stopped so training has enough memory
3. Monitor utilization with nvidia-smi during training

Do not invent local package pins or container tags from this playbook — use the Unsloth docs for current install steps.

Cleanup (optional)

If you created local environments, checkpoints, or downloaded models while following linked Unsloth or Nemotron guidance:

1. Stop any training or notebook session you started
2. Deactivate or remove environments you no longer need
3. Delete downloaded models, datasets, or checkpoints you do not want to keep

Cleanup is optional.

Next steps

• Review Unsloth Documentation for current configuration, install, and hyperparameter guidance
• Read the NVIDIA technical blog on Unsloth and Blackwell
• Explore Nemotron 3 Nano on Hugging Face