Skip to navigation
DocumentationFine-tune Models

Fine-Tuning Tutorials

View as Markdown

Use the tutorials in this section to gain a deeper understanding of how the NVIDIA NeMo Customizer microservice enables fine-tuning tasks.

Tutorials are organized by complexity and typically build on one another. The tutorials reference NHX_BASE_URL, which is the base URL of your NeMo Helix deployment. Refer to the Setup guide for installation, setup, and platform URL guidance.

Getting Started

Understanding Model Entities and Adapters

Learn the fundamentals of how NeMo Customizer works with Model Entities and Adapters, and how to choose the right approach for your project.

model-entities adapters training-types

Dataset Preparation

Format Training Datasets

Learn how to format datasets for different model types.

datasets chat-models completion-models
GRPO Environment Packages

Package native-v1, wheels-v1, and adapter-wheels-v1 environment FileSets for GRPO.

grpo nemo-gym filesets

Customization Jobs

Fine-Tune a Model with Custom Data Using LoRA

Learn how to perform supervised fine-tuning with LoRA adapters using custom data.

nemo-customizer
Fine-Tune a Model with Custom Data Processing All Weights

Learn how to perform supervised fine-tuning using custom data by modifying all training parameters.

nemo-customizer
Align a Model with Direct Preference Optimization

Learn how to run a DPO customization job with preference pairs on the RL backend.

nemo-customizer dpo
Distill a Large Model into a Smaller One with Knowledge Distillation

Learn how to compress a larger teacher model into a smaller student model using knowledge distillation.

nemo-customizer knowledge-distillation
Fine-Tune Nemotron 3 Embed for Domain Retrieval

Mine hard negatives from NVIDIA’s published NVDocs dataset, fine-tune Nemotron 3 Embed 1B, then deploy Retriever NIM and compare nDCG and Recall against the base model.

embedding-models bi-encoder nemo-customizer
Train on SPECTER or Triplet JSONL

Register an existing triplet dataset as a fileset and train on it instead of mined NVDocs data.

embedding-models datasets nemo-customizer

Monitoring & Optimization

Check Customization Job Metrics

Learn how to check job metrics using MLflow or Weights & Biases.

nemo-customizer mlflow wandb
Optimize Tokens per GPU

Learn how to optimize the token-per-GPU throughput for a LoRA optimization job.

nemo-customizer wandb sequence-packing