The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches the official NVIDIA TAO Toolkit container image from nvcr.io and downloads pretrained model weights from HuggingFace (nvidia/Cosmos-Embed1-*). It also references standard model components like google-bert/bert-base-uncased from HuggingFace. All sources are from the official vendor or well-known technology organizations.
[COMMAND_EXECUTION]: The skill executes docker run commands to perform training, evaluation, inference, and model export. These commands utilize specific performance flags such as --gpus all, --ipc=host, and custom ulimit settings, which are standard for high-performance machine learning tasks using the NVIDIA TAO framework.
[CREDENTIALS_UNSAFE]: The skill requires a HuggingFace token (HF_TOKEN) for accessing certain pretrained weights. The documentation explicitly advises against hardcoding the token in configuration files or logs, instructing the user to pass it as an environment variable to the container, which is a secure practice.
[INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided text captions and video data for retrieval and classification tasks. This represents a potential indirect prompt injection surface where malicious instructions could be embedded in data processed by the model.
Ingestion points: User-supplied strings in inference.query.input_texts and metadata JSON files (msrvtt_test_1k.json).
Boundary markers: Data is isolated within structured YAML and JSON configuration files.
Capability inventory: The skill uses Docker to run the cosmos-embed1 CLI for model processing; it does not perform arbitrary shell execution on user-provided strings.
Sanitization: Standard input processing within the TAO Toolkit framework is assumed, though not explicitly detailed in the markdown instructions.

tao-finetune-cosmos-embed