nemotron-customize

nemotron-customize is an NVIDIA skill that plans, configures, and chains repo-native Nemotron model-customization steps into single-step or multi-step pipelines. It covers data curation, translation, SFT and PEFT (via AutoModel or Megatron-Bridge), pretraining and continued pretraining, RL alignment (DPO, RLVR, GRPO, RLHF), BYOB and MCQ benchmark generation, checkpoint conversion, ModelOpt optimization, environment profiles, and evaluation of trained checkpoints or existing hosted endpoints. It solves the problem of translating a model-customization request into a validated, artifact-compatible DAG of existing steps, producing only the YAML/config files needed to run them.

The skill routes primarily from bundled reference files — a step catalog, an artifact-compatibility graph, command references, hardware heuristics, and context packs — and treats live repo files under src/nemotron/steps as the verification and fallback source for exact config fields, runner imports, and manifests. It is deliberately conservative: it does not invent new catalog steps, does not generate new Python or shell outside an approved Explorer mode, and does not guess concrete values such as data paths, model IDs, GPU counts, or auth variable names, returning a Blocked handoff or asking the user instead. Prerequisites include a Nemotron repo checkout, the uv tool to invoke steps, and, for remote or hosted work, an environment profile TOML and the expected auth environment variable (for example NVIDIA_API_KEY), which the guidance insists must stay in the environment and never be inlined or committed.

The skill targets ML engineers and researchers building Nemotron customization and training pipelines. Its handling of credentials is exemplary — official env-var-based auth with explicit no-commit, no-inline guidance — and it runs repo-native training/eval CLI commands rather than arbitrary or remote code, with the user supplying all concrete execution values.