redhat-et/physical-ai-skills

datasets

Use when inspecting, validating, or pulling a user-supplied dataset for fine-tuning or general use, or when asked what data a robot-policy model (pi0.5, DreamZero) was trained or fine-tuned on.

deploy-checkpoint

Use when standing up a finished fine-tuning checkpoint as a live model endpoint for testing or comparison against the base model.

deploy-model

Use when sizing hardware or drafting manifests for a model that fits an existing runtime pattern (stock vLLM chat, or vLLM-Omni multimodal). For a model needing a runtime this platform hasn't run before, use new-model-runtime instead.

fine-tuning

Use when discussing, submitting, or checking on a fine-tuning run — GPU training jobs on pi05/robot-policy checkpoints, submit_finetune_run, get_finetune_run_status, or questions like "how's my fine-tune doing" or "kick off training on X dataset."

model-specs

Use when asked about a catalog model's own static characteristics — architecture, dataset/training compatibility, serving runtime requirements — e.g. "what was DreamZero trained on" or "does pi0.5 need a fixed camera count." Not for live status, scaling, or…

models

Use when listing deployed models, checking a specific model's status/output_kind, scaling it up/down, tearing down a checkpoint deployment, or permanently removing a catalog model.

new-model-runtime

Use when onboarding a model whose serving runtime this platform hasn't run before (not stock vLLM chat or vLLM-Omni) — its own native server/CLI, not something generate_model_manifests can template. See also the deploy-model skill for models that DO fit an…