The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill handles user-provided datasets including images and captions, which represent an ingestion point for external, potentially untrusted data.
Ingestion points: Data is ingested via the paths defined in dataset.train.datasets, dataset.val.datasets, and inference.text_file as seen in SKILL.md and references/skill_info.yaml.
Boundary markers: The instructions do not define specific delimiters or "ignore" instructions for the agent when processing these inputs.
Capability inventory: The skill facilitates the execution of model-related shell commands (clip train, clip evaluate, clip inference, etc.) within a Docker container environment managed by the TAO SDK.
Sanitization: There are no documented sanitization steps for the input text data within the skill's instruction set.
[SAFE]: The skill uses official NVIDIA container images from nvcr.io and refers to internal NVIDIA infrastructure, consistent with its authorship by NVIDIA Corporation. It correctly guides the user on safe secret management by advising against hardcoding credentials and recommending the use of environment-based configuration. It also provides specific security guidance regarding PyTorch checkpoint loading, warning against disabling safety features for untrusted data.

tao-finetune-clip