The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches official TAO Toolkit container images from NVIDIA's container registry (nvcr.io) and retrieves pre-trained model backbones from the Hugging Face model hub (timm library). These are well-known and trusted sources within the machine learning community.
[DYNAMIC_EXECUTION]: Configuration templates (references/spec_template_train.yaml) use a ${eval: ...} syntax for dynamic parameter calculation. This is a built-in feature of the TAO Toolkit used to scale hyperparameters like learning rates based on the total batch size and hardware configuration. These calculations are performed using fixed formulas and are not influenced by untrusted external code.
[INDIRECT_PROMPT_INJECTION]: The skill processes external image datasets for training. The risk of prompt injection through these datasets is mitigated by the nature of the task (vision model training) and the use of Docker volume mounts which scope the file system access to specific, restricted directories.
[SAFE]: The skill follows professional standards for AI model training, including the use of isolated environments, read-only data mounts, and comprehensive JSON schemas to validate all user-provided configuration parameters.

tao-train-dinov3