The Agent Skills Directory

[SAFE]: The skill defines workflows for training and deploying OCRNet models using the NVIDIA TAO toolkit, with all components originating from the trusted NVIDIA registry (nvcr.io).
[SAFE]: Command execution is restricted to legitimate toolkit operations via the Bash tool to manage Docker containers.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external data (character lists, ground truth files, and images), which is a standard surface for model training skills. Evidence chain:
Ingestion points: dataset.character_list_file, dataset_convert.gt_file, and image directories (SKILL.md, skill_info.yaml).
Boundary markers: Not present.
Capability inventory: Shell execution via Bash tool for TAO toolkit commands.
Sanitization: Not present; data is handled by core toolkit binaries.
[DYNAMIC_EXECUTION]: The skill mentions using TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 to load checkpoints in PyTorch 2.6+. This is a specific configuration for trusted development artifacts and does not pose a general risk in this context.

tao-train-ocrnet