Getting Started With Steps#

This guide walks through the basic CLI operations that you use to discover steps, inspect their inputs and outputs, and chain them into pipelines. You do not need a cluster or a GPU to follow along. Every command in this guide reads metadata only.

If you have not read Nemotron Steps Basics, start there for the definitions of step, configuration, environment profile, and artifact.

Prerequisites#

  • Python 3.10 or later.

  • uv installed and on your path.

These prerequisites cover the basics for an introduction to using the steps. For each activity, such as model training or data generation, refer to the getting started page of each activity for additional prerequisites.

Getting Access to the Nemotron CLI#

  1. Clone the repository, if you haven’t already:

    $ git clone https://github.com/NVIDIA-NeMo/Nemotron && cd Nemotron

    Run all commands from the repository root.

  2. Synchronize the common dependencies:

    $ uv sync

  3. Confirm you can run the CLI:

    $ uv run nemotron steps --help

The output lists the step-catalog subcommands that this guide uses: list, show, and run.

List the Available Steps#

Use nemotron steps list to see every step that the CLI discovers. The output is a table with the step identifier, the category, the consumed artifact types, the produced artifact types, and a short description.

$ uv run nemotron steps list

Each row in the table is one building block that you can run, inspect, or chain into a pipeline.

Filter by Category#

Pass --category to narrow the list to one family of steps. For example, the following command shows only the supervised fine-tuning steps.

$ uv run nemotron steps list --category sft

The category matches the top-level folder under src/nemotron/steps/. Common categories include sft, peft, rl, pretrain, optimize, convert, data_prep, sdg, translate, curate, byob, eval, and env.

Emit Machine-Readable JSON#

Pass --json to emit a JSON array instead of a table. The JSON output is the same data the table renders, with no formatting noise, so it works well with jq and with agent tooling.

$ uv run nemotron steps list --json
$ uv run nemotron steps list --category sft --json | jq '.[].id'

The --json flag is available on both nemotron steps list and nemotron steps show.

Inspect a Single Step#

Use nemotron steps show <id> to print the manifest for one step. The output includes the description, the consumed and produced artifacts with their types and descriptions, the parameters with their defaults and choices, and the run specification, which covers the launcher, the container image, the resource shape, and the default configuration.

$ uv run nemotron steps show sft/automodel

The same command in JSON form is the easiest way to read the full manifest programmatically.

$ uv run nemotron steps show sft/automodel --json

The JSON document includes a consumes array and a produces array. Each entry has a type, a required flag, and a description. These two arrays are the contract that connects one step to the next.

Chain Steps Through Artifacts#

Steps fit together because their consumed and produced artifact types match. Use the --produces and --consumes filters on nemotron steps list to walk the building-block graph from either direction.

To find every step that writes a training_jsonl artifact, run the following command.

$ uv run nemotron steps list --produces training_jsonl

The output lists the synthetic data generation and data preparation steps that emit training JSONL.

To find every step that reads a training_jsonl artifact, run the following command.

$ uv run nemotron steps list --consumes training_jsonl

The output lists the fine-tuning and reinforcement-learning (RL) steps that take training JSONL as input.

Together, the two queries show you which steps you can chain. For example, the sdg/data_designer step produces training_jsonl, and the sft/automodel step consumes training_jsonl, so you can pipe the synthetic dataset directly into supervised fine-tuning without an intermediate conversion.

The same pattern works for other artifact types, such as packed_parquet, binidx, checkpoint_hf, checkpoint_megatron, checkpoint_lora, synthetic_jsonl, eval_results, and mcq_benchmark_parquet.

Where To Go Next#

Pick the domain section that matches the work you have in mind.

  • Synthetic Data Generation to generate chat data, tool-calling data, or preference pairs.

  • Translation to translate corpora with optional faithfulness, accuracy, integrity, and translation-quality holistic (FAITH) scoring.

  • Build MCQ Benchmarks to generate a multiple-choice question benchmark from your own documents.

  • Model Training to fine-tune, align, or optimize a model.

  • Model Evaluation to score a deployed checkpoint with standard benchmarks.