Cdf Auto Labeling

NVIDIAOfficialCreated by NVIDIA

Skill
Workflow
Automation and integrations
Apache-2.0
Medium risk

From the author

Use when a user needs to get started with CDF Auto-Labeling, plan a scenario, run or debug a shipped cookbook, author prompts or cookbooks, migrate a pipeline, configure a stage, or pull the prebuilt stage containers from NGC (nvcr.io). Confirm critical inputs (data path, output path, endpoints) and ask when any are missing. This is a router: read the matching reference instead of inventing a workflow.

— the description by NVIDIA

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

CDF Auto-Labeling is a central routing skill designed to guide users through the lifecycle of automated data labeling tasks within the Cosmos Data Factory ecosystem. It provides a structured approach to onboarding, scenario planning, cookbook authoring, and pipeline execution. By acting as a router, it directs users to specialized documentation for specific tasks—such as migrating pipelines, debugging workflows, or managing containerized stages—while enforcing strict input validation and operational guardrails to ensure reproducible results.

What it does

Routes users to specific documentation based on their current task or stage in the pipeline.
Validates critical run inputs like data paths, output directories, and model endpoints.
Provides a standardized workflow for running, debugging, and customizing annotation cookbooks.
Facilitates the integration of new stages and the management of prebuilt container images.
Enforces operational guardrails to prevent common errors like hardcoding secrets or skipping validation steps.

When to use it

When you are a new user needing to get started with CDF Auto-Labeling.
When you need to plan a new annotation scenario or adapt an existing cookbook.
When you are troubleshooting a failed pipeline run or debugging a specific stage.
When you need to migrate an external annotation repository into the CDF framework.
When you need to pull or configure prebuilt stage containers from NGC.

Advantages

Ensures consistent and reproducible pipeline execution through standardized cookbooks.
Prevents configuration errors by mandating input validation before execution.
Simplifies complex workflows by breaking them down into manageable, modular tasks.
Promotes best practices by delegating specialized tasks to dedicated reference guides.

Usage examples

“I cloned the repo and have my own warehouse-safety video. How do I produce auto-labels for my domain?”: The agent will guide you to verify your environment, run a shipped example to confirm the stack, plan your scenario, adapt a cookbook, and execute a dry-run.
“How do I pull the prebuilt stage containers from NGC?”: The agent will direct you to the ngc-containers reference and provide instructions on using make pull to retrieve the necessary images.

Install

Download the skill as a ZIP:
https://autogineer.com/agent-skills/nvidia/cdf-auto-labeling/download

Unzip it into the cross-client skills folder of your project:

mkdir -p .agents/skills && unzip cdf-auto-labeling.zip -d .agents/skills

Documentation by NVIDIA

CDF Auto-Labeling

Use this skill when a user wants to kick off CDF Auto-Labeling on their own data, domain, or use case, or when the request matches a shipped cookbook, stage, authoring, or migration task. This is a router: sequence the specialized references instead of duplicating their detail.

Routing (Read First)

Request looks like	Read
New user, clean checkout, first validated run, "how do I get started"	This file, then the matching reference below
Choose annotation targets / stage subset for a domain	`references/scenario-planning.md`
Create, review, or adapt a cookbook	`references/cookbook-authoring.md`
Write or adapt VLM/LLM prompts or question banks	`references/prompt-authoring.md`
Migrate an existing annotation repo into this one	`references/pipeline-migration.md`
Run the video data augmentation cookbook	`references/video-data-augmentation.md`
Run or choose an EPAS / PAS cookbook	`references/event-and-person-attribute-search.md`
Run event-verification reasoning	`references/event-verification-reasoning.md`
Debug an already-integrated workflow	`references/workflow-runner-debugging.md`
Pull prebuilt stage images from NGC, log in to `nvcr.io`, or fix a missing `cdf-*` image	`references/ngc-containers.md`
Implement or review a new stage or Dockerized service	`references/workflow-stage-integration.md`
Configure or debug one production stage	The matching file under `references/stages/`

Stage references: super-resolution, detection-and-tracking, captioning, visual-qa, reasoning, person-attribute-search, grounding-2d, referring-expressions, training-export.

Instructions

Confirm the critical run inputs with the user before doing anything else, and ask a concise question whenever one is missing or ambiguous - never guess or silently invent a default. At minimum confirm: input data path, output path, VLM/LLM endpoint URLs and model names, model cache path, GPU ids, and (for reasoning-capable models) the max_tokens cap. Restate the confirmed values back to the user before the first execution.
Verify the environment: repository cloned, make targets available, the model cache path exists, the VLM/LLM endpoints are reachable, and a GPU is available. State any missing prerequisite as a blocker instead of assuming it. If the stage images are not present locally, offer to pull the prebuilt images from NGC with make pull instead of building them. Use ngc-containers.
Run a shipped example first to confirm the stack works end to end before customizing. Pick the closest operator pipeline - video data augmentation, event-and-person-attribute-search, or event-verification-reasoning - and run its committed cookbook. Use the matching operator reference.
Plan the target scenario: define modality, domain, intended consumer, and required annotations, and get a minimal stage subset. Use scenario-planning.
Adapt the closest shipped cookbook to the new domain rather than authoring from scratch. Use cookbook-authoring.
Author the domain prompts and question banks. Use prompt-authoring.
Configure the per-stage settings for the domain (detector classes or SAM3 prompts, endpoints, windowing, max_tokens). Use the relevant stage reference, starting with detection-and-tracking.
Dry-run the adapted cookbook, then execute and validate the outputs. Use workflow-runner-debugging.

Adopting an existing external annotation or dataset-generation repository into CDF instead of starting from a shipped cookbook is a migration task; use pipeline-migration for that path.

Examples

New user, new domain: "I cloned the repo and have my own warehouse-safety video. How do I produce auto-labels for my domain?"

Guided path:

Confirm env (model cache, VLM/LLM endpoints, GPU), then prove the stack on a shipped example before customizing:

make run SCRIPT=workflow-runner:main \
  ARGS='--cookbook-file cookbooks/video_data_augmentation/configs/pipeline_video.yaml --container-dry-run'

Plan the domain (scenario-planning) -> subset detection_and_tracking -> captioning -> visual_qa -> reasoning -> training_export (add grounding_2d for caption→boxes or referring_expressions for boxes→phrases; use grounding-2d / referring-expressions).
Copy the closest cookbook to cookbooks/warehouse_safety/configs/pipeline.yaml and adapt inputs, detector classes/SAM3 prompts, prompts, and question banks.
Dry-run the new cookbook, then run for real and validate outputs:

make run SCRIPT=workflow-runner:main \
  ARGS='--cookbook-file cookbooks/warehouse_safety/configs/pipeline.yaml --container-dry-run'

Guardrails

Do not guess or fabricate the critical inputs enumerated in step 1; if any is missing or ambiguous, ask the user and confirm before executing.
Do not customize a cookbook before a shipped example runs clean; a broken base makes domain debugging ambiguous.
Keep the first custom pipeline minimal - only the stages needed for the requested annotations - and expand later.
Verify that every selected stage's service package and image exist in the current branch before promising an end-to-end run.
Do not put secrets, tokens, or absolute home paths in committed cookbooks; use placeholders such as <model-cache> and env vars for endpoint keys.
For reasoning-capable models (for example Gemini 3 Flash), raise max_tokens on the visual_qa and reasoning LLM substages to avoid the thinking-token tax; keep the default cap for non-reasoning models.
Do not rely on non-CDF pipelines, commands, or file locations. A first run must be reproducible through workflow-runner:main inside this repo.

Files

ATTRIBUTION.mdAdded by AutoGineer909 B
BENCHMARK.md7.3 KiB
LICENSELicence file added by AutoGineer11.2 KiB
SKILL.md7.5 KiB
evals/evals.json5.2 KiB
references/README.md1.3 KiB
references/cookbook-authoring.md6.3 KiB
references/cookbook-authoring/cookbook-schema.md6.3 KiB
references/cookbook-authoring/scenario-authoring.md2.8 KiB
references/event-and-person-attribute-search.md5.1 KiB
references/event-and-person-attribute-search/run-reference.md2.4 KiB
references/event-verification-reasoning.md1.9 KiB
references/event-verification-reasoning/run-reference.md1.3 KiB
references/ngc-containers.md4.2 KiB
references/pipeline-migration.md7.2 KiB
references/pipeline-migration/migration-playbook.md3.8 KiB
references/pipeline-migration/nonlinear-video-pipeline-example.md2.7 KiB
references/pipeline-migration/service-reuse-matrix.md3.4 KiB
references/pipeline-migration/sidecar-contracts.md3.0 KiB
references/prompt-authoring.md3.3 KiB
references/prompt-authoring/domain-evidence-checklists.md2.2 KiB
references/prompt-authoring/prompt-patterns.md3.9 KiB
references/prompt-authoring/question-bank-authoring.md2.0 KiB
references/scenario-planning.md3.6 KiB
references/scenario-planning/annotation-recommendations.md5.0 KiB
references/scenario-planning/intake-and-decisions.md2.6 KiB
references/scenario-planning/pipeline-composition.md6.0 KiB
references/stages/captioning.md3.6 KiB
references/stages/detection-and-tracking.md6.2 KiB
references/stages/detection-tracking-patterns.md10.3 KiB
references/stages/grounding-2d-patterns.md1.6 KiB
references/stages/grounding-2d.md5.1 KiB
references/stages/guardrails.md532 B
references/stages/person-attribute-search.md4.5 KiB
references/stages/reasoning.md5.0 KiB
references/stages/referring-expressions-patterns.md1.2 KiB
references/stages/referring-expressions.md5.5 KiB
references/stages/super-resolution.md4.4 KiB
references/stages/training-export.md3.7 KiB
references/stages/visual-qa.md6.2 KiB
references/video-data-augmentation.md12.2 KiB
references/video-data-augmentation/run-reference.md3.7 KiB
references/workflow-runner-debugging.md6.0 KiB
references/workflow-runner-debugging/data-handoff.md3.5 KiB
references/workflow-runner-debugging/runner-debugging-patterns.md3.7 KiB
references/workflow-stage-integration.md7.1 KiB
references/workflow-stage-integration/stage-service-contract.md3.0 KiB
references/workflow-stage-integration/validation-checklist.md2.1 KiB
skill-card.md4.9 KiB
skill.oms.sig15.6 KiB

references/scenario-planning/annotation-recommendations.md

Annotation Recommendations

Use this matrix to recommend annotation outputs. Keep the recommendation tied to modality, domain, and downstream consumer.

By Modality

Image

Usually useful:

Image caption: visible scene/object/activity evidence.
VQA evidence: question-bank answers for domain-specific attributes.
reasoning: when downstream expects DAFT task//contextual files.
training_export: when outputs feed a training/export dataset.
person_attribute_search: for per-person attribute search labels.
Detection: only if object boxes are needed; skip for scene-level captions.

Avoid by default:

Temporal localization, events, causal linkage, and video summaries unless the image task explicitly defines a non-temporal equivalent.
Super-resolution unless the task needs improved readability and the SR service supports the input modality.

Video

Usually useful:

Dense window captions: temporal visual evidence and activity summaries.
Detection/tracking: if object identity, counts, trajectories, overlays, masks, or object-level grounding are needed.
VQA evidence: if the task has reusable questions or decision criteria.
reasoning: if final artifacts should be DAFT task//contextual outputs.
training_export: if per-scene outputs should be aggregated into a dataset.
Temporal localization: if answers need start/end times and services in the current branch support it.
Causal linkage: only when event chains are visible enough to support it and services in the current branch support it.

Optional:

Super-resolution for low-resolution clips before downstream stages.
SAM3 when promptable masks/tracks or annotated videos are needed.

Detector Choice

Choose RF-DETR when COCO boxes and stable track IDs are enough. It is the default for traffic/person/vehicle inventory and is cheaper than SAM3.
Choose SAM3 when the target objects are not COCO classes, when masks/contours are required, or when a VLM should reason over an annotated ID video.
Request tracking overlays only when downstream prompts need visible IDs. For captioning to consume those overlays, set --input-source tracking or rely on auto only after verifying the overlay path exists.
Do not let detector limitations define the annotation goal. Use VLM/VQA for open-vocabulary evidence and use reasoning/training_export for final schemas and dataset aggregation.

By Domain

Traffic / Roadway Safety

Recommended outputs: dense captions, vehicle/person tracks, event summaries, temporal localization when supported by services in the current branch, VQA for roadway conditions and incidents, reasoning verdicts, and training_export. Use SR when clips are low resolution. Use RF-DETR for COCO traffic classes. Use SAM3 for promptable incident actors, road debris, emergency vehicles, masks, or annotated ID videos used by captioning/VQA.

Warehouse / Operational Liability

Recommended outputs: dense captions, worker/equipment tracks when available, PPE/hazard VQA, temporal localization for near-misses/incidents when supported by services in the current branch, reasoning, and training_export. Use RF-DETR for person and coarse COCO proxies such as truck; use SAM3 for forklifts, pallets, pallet jacks, boxes, tools, and other promptable objects.

Security Surveillance

Recommended outputs: person tracks, dense captions, VQA for entry/exit, loitering, object handling, confrontation, visible weapons, and response cues. Avoid identity or demographic inference. Use RF-DETR for person/vehicle/bag COCO boxes; use SAM3 only for concrete visible promptable objects, not intent labels.

Employee Conduct

Recommended outputs: person tracks when full-frame video is used, dense captions, VQA for on-task/off-task activity, station presence, PPE/badge/uniform visibility, and interaction tone. Avoid intent and employment-status inference. Use SAM3 only when tool/workcell object IDs or masks materially improve evidence.

Robotics

Recommended outputs: dense captions, task-phase VQA, success/failure evidence, object/tool interaction evidence, safety proximity, and optional SAM3 prompts for specific objects/tools. Avoid inferring robot intent or hidden success criteria. Use RF-DETR for people and COCO objects; use SAM3 for robot arms, grippers, workpieces, tools, hands, and task-specific promptable objects.

Person Attributes

Recommended outputs: image caption and VQA/open QA about visible clothing, accessories, carried items, pose, crop quality, and occlusion. For the Event and Attribute Search product, run detection_and_tracking then person_attribute_search for per-track attributes. Avoid protected or identity attributes.

Generic Image Annotation

Recommended outputs: image caption, scene-level VQA, optional detection for object inventory, and reasoning/training_export when required by downstream systems.

Minimality Rule

Recommend the smallest stage set that produces the requested annotations. Add expensive stages only when they improve the evidence needed by downstream tasks.

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Security

What our scanner noticed· Medium risk

Large base64-encoded blobs

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by

NVIDIA

Publisher

NVIDIA (Official)

Source

NVIDIA/skills · skills/cdf-auto-labeling · commit fce3cca

Licence

Apache License 2.0 (redistribution allowed)

LICENSE (added by AutoGineer from LICENSE-APACHE)
ATTRIBUTION.md (written by AutoGineer)

Synced

2026-11-09

AutoGineer is not affiliated with or endorsed by NVIDIA or NVIDIA.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download cdf-auto-labeling.zip