nemotron-customize

To plan Nemotron customization, sequence curated JSONL translation, SFT/PEFT, pretraining CPT, RL alignment, and BYOB benchmarks into cohesive pipelines. The workflow routes goals through bundled catalogs, deduces artifact DAGs, and generates validated per-step YAML configs.

Generating YAML configs for Nemotron SFT and RL alignment involves loading live step manifests from the catalog and producing new config files under src/nemotron/steps/<cat>/<step>/config/. The process ensures artifact compatibility and validates prerequisites before execution.

Yes, Explorer mode supports plan-first workflows for Nemotron customization when no exact catalog surface exists. It enables safe scoping and planning of data curation, training, and evaluation steps before generating any actual YAML configuration files or code changes.

Yes, Nemotron customization pipelines support checkpoint conversion and model optimization. The workflow orchestrates these stages alongside pretraining CPT, RL alignment, and BYOB benchmarks, ensuring artifact compatibility across all stages using the official catalog and artifact graphs.

Artifact graphs ensure compatibility across Nemotron training stages by deducing artifact DAGs that map dependencies between data curation, translation, SFT/PEFT, and RL alignment. The pipeline validates prerequisites against bundled ARTIFACTS and CATALOG references before generating step configs.

Nemotron customization relies on bundled ARTIFACTS, CATALOG, PATTERNS, and HARDWARE references to route tasks. If a requested pipeline lacks an exact catalog surface, users must switch to Explorer mode for safe scoping, as standard routing cannot deduce artifact DAGs for unmapped configurations.