The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for executing the NVIDIA TAO Data Services workflow using Python modules and Docker commands. These operations are intended for the skill's primary purpose of model training and data mining and are performed within a containerized environment to isolate the host.
[INDIRECT_PROMPT_INJECTION]: The skill ingests unlabeled image pools and data manifests for mining and selection. While this creates a processing surface for external data, the risk is addressed through the following architecture:
Ingestion points: Data enters the workflow via Parquet manifests (data.target_manifest) and source payload contracts (data.source_payload_contract).
Boundary markers: The skill requires the agent to run validate and preflight commands to verify cohort sizes, schema compliance, and configuration before initiating the loop.
Capability inventory: The skill uses Read, Write, and Bash tools to manage configuration files and launch the training controller; training itself is handled by specialized subprocesses (dinov3 train).
Sanitization: The controller uses strict JSON schema validation for workflow configurations and metrics, and implements materialization seals (SHA-256 digests) to ensure data lineage and prevent accidental contamination from modified files.
[DYNAMIC_EXECUTION]: The workflow allows for user-supplied custom scoring or evaluation logic via implementation_files. This is a documented extensibility feature, and the skill provides a "path contract" and approval workflow to ensure all dependencies are identified and approved before the run starts.

tao-run-dinov3-ssl-deft