Nemotron Customization Pipelines
Plan, configure, and chain existing Nemotron steps for data preparation, training, alignment, conversion, optimization, and evaluation.
The skill limits actions to repository-safe commands, prohibits environment and secret disclosure, requires auth values to remain external, and calls for confirmation before remote, destructive, or expensive execution. It also exposes some data-flow and hosted-service boundaries. Deducted 5 points because it lacks a complete sensitive-data classification, general rollback procedure, and post-execution recovery plan.
The catalog, artifact graph, command reference, hardware guidance, and troubleshooting rules are broadly self-consistent and include blocking conditions and diagnosable feedback; BENCHMARK.md reports NVSkills-Eval results. Deducted 12 points because this review did not execute key paths, no committed test suite covering the skill's core routes is provided, and missing live repository or configuration details may still shift verification work to the user.
The name, triggers, intended scenarios, and exclusions are reasonably clear, with distinctions among AutoModel, Megatron-Bridge, and BYOB workflows. Deducted 6 points because Chinese-user support and mainland-China network reachability are not addressed, some hosted workflows depend on external authentication and network access, and trigger precision lacks validated negative cases.
The documentation has strong information architecture with purpose, prerequisites, limits, routing, references, examples, configuration constraints, troubleshooting, version, Apache-2.0 licensing, and authorship; the README describes repository mirroring and updates. Deducted 5 points for an overly long description, no explicit changelog, unclear skill-level maintenance ownership or upgrade procedure, and exact commands that still require live verification.
The skill maps common Nemotron customization requests to existing steps and artifact types and provides usable command shapes, required inputs, and safety callouts; BENCHMARK.md reports strong correctness and effectiveness results on core tasks. Deducted 9 points because static review cannot verify command execution or output quality, and the benchmark gives only aggregate results with no negative tasks or raw execution evidence, so direct usability is conservatively limited.
The skill cites traceable repository-local catalogs, artifact rules, command references, and patterns, and includes a committed evaluation report with eight task records. Deducted 6 points because there was no independent execution, CI evidence, complete test artifacts, or cross-source corroboration; the reported evaluation is primarily self-described and coverage is limited.
- Before execution, verify the CLI, configuration fields, environment profile, and dependency versions in the target Nemotron checkout; the skill itself does not prove current runnability.
- Translation and hosted evaluation may send user data to external services and depend on authentication and network reachability; mainland-China users should confirm endpoint access and compliance first.
What does this skill do, and when should you use it?
This skill supports end-to-end Nemotron model-customization workflows. It can combine curation, translation, SFT/PEFT, pretraining, RL alignment, benchmarking, checkpoint conversion, ModelOpt optimization, and evaluation into single-step or multi-step pipelines. It plans the artifact DAG, validates wiring and configuration, and creates YAML/config files for existing repository steps. It is not intended for generic ML advice, deployment-only work, frontend or visualization tasks, or unrelated coding.
Reads the Nemotron step catalog, bundled references, artifact definitions, command guidance, hardware guidance, selected live step files, and configuration files; routes requests to existing steps; plans artifact dependencies for pipelines; validates selected configurations and environment profiles; creates the required YAML/config files; prepares command shapes using uv run nemotron steps ... for local runs, dry-runs, or remote submissions; and identifies required authentication environment variables for translation or hosted evaluation without exposing their values.
- A Nemotron developer needs to curate or translate data before connecting it to a training step.
- A user needs SFT or PEFT for a Hugging Face model and must choose between AutoModel and Megatron-Bridge.
- A team needs to compose data preparation, training, alignment, checkpoint conversion, and evaluation into a multi-stage pipeline.
- A user needs to benchmark or smoke-test an existing checkpoint, API model ID, deployed model, or hosted endpoint.
- A model team needs to apply DPO, RLVR, GRPO, or RLHF alignment to a Nemotron workflow.
What are this skill's strengths and limitations?
- Covers a broad Nemotron workflow from data preparation through training, alignment, conversion, optimization, and evaluation.
- Prioritizes existing repository steps and validates artifact edges instead of inventing catalog capabilities.
- Includes concrete guardrails for configuration scope, secret handling, checkpoint selection, and training-parameter alignment.
- Accounts for local, dry-run, remote-submission, and hosted-service workflows.
- Requires a Nemotron checkout containing src/nemotron/steps/ and an available uv command.
- Will not guess model IDs, paths, GPU counts, profiles, or task IDs, so incomplete inputs block runnable output.
- Requires planning and approval before creating configuration for multi-stage pipelines.
- The supplied materials provide no skill-specific test suite, supported-platform matrix, or complete per-step execution reference.
- Does not cover deployment-only tasks, frontend or dashboard work, visualization, generic ML advice, or non-Nemotron tasks.
How do you install this skill?
Install the specific skill using the repository README's supported skills CLI command:
npx skills add nvidia/skills --skill nemotron-customize --yes
The provided README does not specify a fixed installation directory; the CLI prompts for the destination.
How do you use this skill?
After installation, mention the relevant Nemotron step or workflow in a task prompt, for example: “Plan a Nemotron pipeline for data curation, SFT, and evaluation.” Single-step prompts such as “evaluate this existing Nemotron checkpoint” or “fine-tune a Nemotron model with PEFT” are also within scope. Before producing runnable commands, provide the repository checkout, concrete model or checkpoint, data paths, output directory, hardware/GPU count, and any required remote profile or authentication setup. The supplied materials do not document the complete per-step command syntax or configuration fields.
How does this skill compare with similar options?
The skill explicitly treats AutoModel and Megatron-Bridge as alternative SFT/PEFT backends and directs users to weigh hardware, GPU count, and workload constraints. The supplied materials do not support a broader competitor comparison.