The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides instructions to set the environment variable TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 to enable loading of Lightning checkpoints in PyTorch 2.6. This disables weights-only loading, permitting the use of pickle.load for complex metadata which can allow for arbitrary code execution if a malicious checkpoint is provided. The skill documentation includes warnings to use this setting only with trusted checkpoints produced by the TAO workflow.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external text data through image captions and inference prompts which are interpolated into model operations.
1. Ingestion points: Image caption directories (dataset.train.datasets) and inference text files (inference.text_file).
1. Boundary markers: No specific delimiters or boundary markers are documented in the spec templates to separate untrusted data from model instructions.
1. Capability inventory: The skill has access to the Bash tool and executes clip CLI actions within Docker containers for training, evaluation, and inference.
1. Sanitization: Optional text canonicalization (model.canonicalize_text) is available for input normalization.

tao-finetune-clip