Model Training How-To Guides

This section provides task-focused guides for common training workflows. For your first run, start with Getting Started with Training Steps.

If you are new to fine-tuning concepts, read Training Concepts for definitions of fine-tuning approaches, data formats, and checkpoints before working through these tasks.

Run SFT with AutoModel on Custom Data

Configure tiny.yaml for your own JSONL data, run the step, verify the checkpoint output, and resolve common issues.

30 min beginner

Run SFT with AutoModel on Custom Data

Choose an SFT Backend

Pick between sft/automodel and sft/megatron_bridge based on checkpoint format and scale.

intermediate

Choose an SFT Backend

Choose a PEFT Backend

Pick between peft/automodel and peft/megatron_bridge based on base checkpoint format and data path.

intermediate

Choose a PEFT Backend

Choose an RL Alignment Step

Pick between DPO, RLVR, and RLHF based on how the reward signal enters training.

intermediate

Choose an RL Alignment Step

Run Post-Training Optimization

Apply quantization, pruning, or distillation with Model Optimizer after a trained model passes your quality bar.

intermediate

Run Post-Training Optimization

Environment Profiles and Executors

Configure NeMo Run environment profiles for local, Slurm, and Lepton execution.

intermediate

Environment Profiles and Executors

Data and Checkpoint Formats

Understand the JSONL, Parquet, and checkpoint types that training steps declare in step.toml.

intermediate

Data and Checkpoint Formats

Convert Checkpoints Between Training Steps

Run a conversion step when one step produces a checkpoint layout that the next step cannot consume directly.

intermediate

Convert Checkpoints Between Training Steps