Training Reference#

This section provides lookup material for every supervised fine-tuning (SFT), parameter-efficient fine-tuning (PEFT), reinforcement learning (RL), and optimization step packaged under src/nemotron/steps/. Use these pages to find the exact CLI syntax, configuration file layout, parameters, and configuration overrides for each step.

Reference Sections#

Nemotron Steps CLI Reference

Shared command-line syntax, options, dotlist overrides, and passthrough arguments for the nemotron steps command group.

lookup

Nemotron Steps CLI Reference
Env Profile Generator

The packaged env/env_toml step that generates the environment profile file every other training step consumes.

setup

Env Profile Generator
Step Catalog

Every step identifier, manifest path, and per-step reference link.

lookup

Step Catalog
Configuration Conventions

Per-step config/ layout, CLI configuration resolution, dotlist override rules, and environment-variable expansion.

lookup

Configuration Conventions

Per-Category Step References#

Supervised Fine-Tuning Steps

The sft/automodel and sft/megatron_bridge references.

reference

Supervised Fine-Tuning Steps
Parameter-Efficient Fine-Tuning Steps

The peft/automodel and peft/megatron_bridge references.

reference

Parameter-Efficient Fine-Tuning Steps
Reinforcement Learning Steps

The rl/nemo_rl/dpo, rl/nemo_rl/rlvr, and rl/nemo_rl/rlhf references.

reference

Reinforcement Learning Steps
Optimization Steps

The optimize/modelopt/quantize, optimize/modelopt/prune, and optimize/modelopt/distill references.

reference

Model Optimization Steps
Checkpoint Conversion Steps

The convert/hf_to_megatron, convert/megatron_to_hf, and convert/merge_lora references.

reference

Checkpoint Conversion Steps