The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides detailed instructions for running Docker commands to perform model training, evaluation, and inference. These commands are standard for NVIDIA TAO workflows and are used as intended for the skill's primary function. The instructions explicitly set TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 to allow loading specific checkpoint formats, which is a documented requirement for this vendor toolchain.
[EXTERNAL_DOWNLOADS]: The skill pulls Docker containers from the official NVIDIA registry (nvcr.io). It also facilitates downloading pre-trained weights from Hugging Face and internal S3 storage. These sources are considered trusted and well-known services within the context of the skill author (NVIDIA).
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted metadata in JSON format (vadr1_chunks) which contains text queries and captions.
Ingestion points: Metadata files specified in dataset.train.video_text.metadata and dataset.val.video_text.metadata in SKILL.md and references/skill_info.yaml.
Boundary markers: Absent.
Capability inventory: Execution of the video_clip CLI via Docker subprocess calls as described in SKILL.md.
Sanitization: Not explicitly defined in the instructions, as data is processed by the underlying machine learning framework. The risk is minimized by the structured nature of the JSON input and the specific use case (model training).

tao-finetune-video-clip