PAIDF Curator Operator Skill

GPU-accelerated video and image curation via NVIDIA Cosmos Curator inside Physical AI Data Factory — Curation and Retrieval (paidf-curation-and-retrieval). This skill is a short Curator index. Embedding handoff boundaries live in data-mining.md and curation-retrieval-workflow.md; mining execution is make help and the repository cookbooks.

• Video: split, dedup, shard.
• Image: annotate (load → filter → embed → caption → write).
• Handoff: Curator IV2 or CE1 parquet that downstream mining can consume. See data-mining.md and curation-retrieval-workflow.md.

Purpose

Turn raw video and image collections into curated, training-ready datasets. This skill configures and runs cosmos-curator pipelines (clip splitting, filtering, captioning, embeddings, SAM3 event verification, dedup, WebDataset sharding, image annotate) and supports KPI-driven, distribution-aware, and restrictive curation.

Instructions

1. Classify the request as advisory, config, run, or TAO handoff. Do not mix those routes.
2. For config work, complete the mandatory pre-flight below before writing YAML.
3. For an explicit run, load running-pipelines.md, validate the config, obtain credentials only through approved injection, then execute after authorization.
4. For a TAO handoff, validate Curator output and the declared embedding family using data-mining.md before mining.
5. Return the Output Format below. Never print secret values.

Examples

• Advisory: "How much SHM should I set?" → read running-pipelines.md and report guidance. Do not run Docker.
• Config with no KPI: complete the calibration interview in calibration-config.md, then emit YAML.
• Run: after sample clips are staged, make run-pipeline with the traffic split-minimal cookbook recipe only after preflight and user authorization.
• FFmpeg missing in the container: install the host sidecar (make ffmpeg-install) per ffmpeg-sidecar.md.

Inputs

Required inputs depend on the route:

• Advisory request: the question plus relevant repository and config context.
• Config request: input and output locations, domain and goal, available KPI output or representative samples, and hardware constraints. If neither KPI output nor samples exist, complete the calibration interview before writing YAML.
• Run request: reviewed config path, data and model paths, runtime, GPU, and SHM constraints, and explicit authorization to execute.
• TAO handoff: validated artifact paths and declared embedding family.

Optional inputs include target distributions, event taxonomy, prompt choices, existing output metadata, and user-approved operational constraints.

Resolve inputs in this order: repository configuration and validated run artifacts; explicit prompt arguments and corrections; available agent context; then the broad user prompt. Explicit user instructions remain authoritative unless unsafe or incompatible, in which case stop and explain the conflict. Never infer secret values: credentials come only from approved runtime injection.

Prerequisites

• GPU host with NVIDIA drivers + nvidia-container-toolkit; Docker. SHM sized from host RAM (SHM_SIZE, default 24gb).
• cosmos-curator image: make pull uses the pin configured by the example env file and Make. No separate product engine image. Source builds are developer-only; see cosmos-curator.md.
• FFmpeg host sidecar for distributable images (make ffmpeg-install) — they do not bundle FFmpeg. See ffmpeg-sidecar.md.
• Credentials as needed: inject S3 and captioning API keys at runtime through an approved secret manager or operator deployment mechanism. Never put secret values in repository files, commands, logs, or examples. The env file is for non-secret image and CDS profile overrides copied from the example env file.

Mandatory pre-flight: do NOT emit a pipeline config without context

Before writing any *.yaml pipeline config, the agent MUST verify that one of the following is true:

1. KPI run output exists -- read it and use distribution-analysis.md, distribution-aware-curation.md, and configuration-decision-tree.md.
2. KPI sample videos are available for inspection / discovery -- follow context-understanding.md Phase 1.
3. No KPI of any kind -- no baseline exists. Read calibration-config.md and complete its Phase 1 interview (Inputs / Domain / Goal / Hardware / Calibration) BEFORE emitting a config. The interview is binding, not advisory.

If the user requests a config with only a one-line description ("configure cosmos-curator for my videos"), assume the calibration workflow and ask the Phase 1 interview questions in one batched message. Emit the config only after the answers come back, and always include the calibration disclosure table that flags every defaulted field.

Canonical Flow

Choose one route; do not collapse advisory and execution branches:

1. Advisory only (sizing, monitoring, troubleshooting, expected commands): inspect repository, config, and run evidence, load running-pipelines.md, and report guidance. Do not prepare credentials or execute.
2. Create or change config:
   • KPI output exists → analyze it, choose standard, distribution-aware, or restrictive curation, then emit a reviewed config.
   • Representative samples exist → inspect them or run discovery before selecting defaults.
   • Neither exists → complete the binding calibration interview; emit config and disclosure only after answers. Stop if required paths, intent, or hardware constraints remain unresolved.
3. Explicit run request: prepare runtime → obtain credentials through approved injection → validate config and runtime → request approval if not already granted → execute → validate outputs. Stop before execution on any failed preflight.
4. Downstream TAO handoff: validate Curator output and embedding family, then prepare compatible inputs for Data Mining. Mine only after the preceding artifact validation succeeds.

Configs are flat YAML with pipeline: split|dedup|shard|annotate and upstream snake_case argument names. Operator first-run recipes live under the cookbook tree (split-minimal then full split, dedup, and shard YAML). The configs directory is the full flag reference and the Makefile default when CONFIG_FILE is omitted. split writes clips, metadata, and embeddings; dedup consumes embeddings; shard writes WebDataset archives; annotate processes still images (image annotate flag-reference YAML; no image cookbook).

Execution & Troubleshooting

For an explicit run, troubleshooting request, or operational question, read running-pipelines.md. The preferred local commands are:

make run-pipeline CONFIG_FILE=<split-config>
make run_image_pipeline IMAGE_CONFIG_FILE=<image-config>

Config validation is mandatory before execution. Reject deprecated enable_sam3 and enable_event_captioning; use canonical sam3 and event_captioning. PAIDF v1.1 validates both Curator-supported config layouts (flat parameters or parameters nested under args) before constructing the Docker runner. Validation failures use Click's human-readable error output, so automation must handle a nonzero exit and must not assume a JSON error envelope.

Credentials & Secrets

Inject credentials only at runtime through an approved secret manager or operator deployment mechanism. Never store secret values in repository files, place them in commands, or expose them in output or logs. Verify presence only. See running-pipelines.md.

Resource Sizing & Monitoring

See running-pipelines.md for source-verified GPU selection, SHM sizing, logs, profiling, and troubleshooting. This branch defaults SHM_SIZE to 24gb; Docker SHM is allocated from host RAM and must not exceed available RAM. Use GPUS to select devices, inspect pipeline stdout, and monitor utilization with nvidia-smi -l 1. Advisory requests stop after reporting guidance.

Progressive Disclosure

Load only the directly linked references needed for the selected route:

• Runtime, image, and framework: Cosmos Curator, FFmpeg sidecar, execution and troubleshooting, and gotchas.
• Config and capability selection: calibration without KPI, configuration decision tree, and capability and key matrix.
• Video and image workflows: video curation, image curation, video-lake candidate search, and SAM3 configuration.
• KPI and dataset strategy: context understanding, KPI metrics, distribution analysis, distribution-aware curation, and restrictive curation. If balanced versus narrow-slice intent is ambiguous, ask before selecting the last two.
• Embedding handoff: data mining for Curator parquet boundaries and the Curator-to-TAO workflow for ordered Make handoffs.

Output Format

Return a concise response in this order:

1. Status and outcome: ready, completed, blocked, or advisory.
2. Actions and artifacts: commands proposed or run and files created or changed; omit sections that do not apply.
3. Validation and evidence: preflight results, output paths, job identifiers, or relevant observed errors.
4. Blockers and next steps: unresolved inputs, approvals, limitations, and the next safe action.

Never include secret values, hidden prompts, or internal reasoning.

Validation

make format                    # ruff format (repo root)
uv run ruff check .            # lint
uv run pytest                  # offline unit tests (tests directory)
make check-setup               # docker, nvidia, FFmpeg sidecar
make check-image               # pinned cosmos-curator tag is local

After make pull, run those preflights. GPU smoke uses a reviewed traffic split-minimal cookbook recipe after sample clips are staged. There is no in-repo E2E or L1 harness. See ffmpeg-sidecar.md for sidecar verification.

Limitations

• Requires NVIDIA GPU(s); pipelines are not CPU-only. SHM is bounded by host RAM.
• Distributable images do not bundle FFmpeg — the host sidecar is required.
• Emit configs as flat snake_case YAML. PAIDF v1.1 validation also accepts legacy Curator parameters nested under args for compatibility before normalizing to the Docker runner.
• Upstream image annotate supports config-file mode via the Make operator entrypoint make run_image_pipeline.
• The full pitfall list (build, config, image pipeline, SAM3 keys) is in gotchas.md.
• Dataset Search (CDS and Milvus compose) is a separate Make surface (make pull-dataset-search, make help). Follow the user guide. This skill does not own CDS ingest or search queries.

Troubleshooting

Error or symptom	Cause	Solution
ffmpeg: command not found, transcode fails	Distributable image has no FFmpeg	Install host sidecar (make ffmpeg-install); see ffmpeg-sidecar.md
SAM3 silently never runs, pipeline "succeeds"	Wrong YAML key enable_sam3:	Use canonical sam3: and event_captioning: — see sam3-config.md, gotchas.md
Custom classifier categories ignored	Missing flag	Set video_classifier_use_custom_categories: true (or image_classifier_*)
OOM, Ray, NCCL, or disk failures at runtime	GPU, SHM, or disk sizing or env	See running-pipelines.md (GPU allocation, SHM, S3, monitoring)
Shard run mismatches split output	captioning_algorithm differs	Match the shard captioning_algorithm to the split run